tools/perf/util/machine.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <dirent.h>
   3 #include <errno.h>
   4 #include <inttypes.h>
   5 #include <regex.h>
   6 #include <stdlib.h>
   7 #include "callchain.h"
   8 #include "debug.h"
   9 #include "dso.h"
  10 #include "env.h"
  11 #include "event.h"
  12 #include "evsel.h"
  13 #include "hist.h"
  14 #include "machine.h"
  15 #include "map.h"
  16 #include "map_symbol.h"
  17 #include "branch.h"
  18 #include "mem-events.h"
  19 #include "path.h"
  20 #include "srcline.h"
  21 #include "symbol.h"
  22 #include "sort.h"
  23 #include "strlist.h"
  24 #include "target.h"
  25 #include "thread.h"
  26 #include "util.h"
  27 #include "vdso.h"
  28 #include <stdbool.h>
  29 #include <sys/types.h>
  30 #include <sys/stat.h>
  31 #include <unistd.h>
  32 #include "unwind.h"
  33 #include "linux/hash.h"
  34 #include "asm/bug.h"
  35 #include "bpf-event.h"
  36 #include <internal/lib.h> // page_size
  37 #include "cgroup.h"
  38 #include "arm64-frame-pointer-unwind-support.h"
  39
  40 #include <linux/ctype.h>
  41 #include <symbol/kallsyms.h>
  42 #include <linux/mman.h>
  43 #include <linux/string.h>
  44 #include <linux/zalloc.h>
  45
  46 static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
  47                                      struct thread *th, bool lock);
  48 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip);
  49
  50 static struct dso *machine__kernel_dso(struct machine *machine)
  51 {
  52         return map__dso(machine->vmlinux_map);
  53 }
  54
  55 static void dsos__init(struct dsos *dsos)
  56 {
  57         INIT_LIST_HEAD(&dsos->head);
  58         dsos->root = RB_ROOT;
  59         init_rwsem(&dsos->lock);
  60 }
  61
  62 static void machine__threads_init(struct machine *machine)
  63 {
  64         int i;
  65
  66         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
  67                 struct threads *threads = &machine->threads[i];
  68                 threads->entries = RB_ROOT_CACHED;
  69                 init_rwsem(&threads->lock);
  70                 threads->nr = 0;
  71                 INIT_LIST_HEAD(&threads->dead);
  72                 threads->last_match = NULL;
  73         }
  74 }
  75
  76 static int thread_rb_node__cmp_tid(const void *key, const struct rb_node *nd)
  77 {
  78         int to_find = (int) *((pid_t *)key);
  79
  80         return to_find - (int)thread__tid(rb_entry(nd, struct thread_rb_node, rb_node)->thread);
  81 }
  82
  83 static struct thread_rb_node *thread_rb_node__find(const struct thread *th,
  84                                                    struct rb_root *tree)
  85 {
  86         pid_t to_find = thread__tid(th);
  87         struct rb_node *nd = rb_find(&to_find, tree, thread_rb_node__cmp_tid);
  88
  89         return rb_entry(nd, struct thread_rb_node, rb_node);
  90 }
  91
  92 static int machine__set_mmap_name(struct machine *machine)
  93 {
  94         if (machine__is_host(machine))
  95                 machine->mmap_name = strdup("[kernel.kallsyms]");
  96         else if (machine__is_default_guest(machine))
  97                 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
  98         else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
  99                           machine->pid) < 0)
 100                 machine->mmap_name = NULL;
 101
 102         return machine->mmap_name ? 0 : -ENOMEM;
 103 }
 104
 105 static void thread__set_guest_comm(struct thread *thread, pid_t pid)
 106 {
 107         char comm[64];
 108
 109         snprintf(comm, sizeof(comm), "[guest/%d]", pid);
 110         thread__set_comm(thread, comm, 0);
 111 }
 112
 113 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
 114 {
 115         int err = -ENOMEM;
 116
 117         memset(machine, 0, sizeof(*machine));
 118         machine->kmaps = maps__new(machine);
 119         if (machine->kmaps == NULL)
 120                 return -ENOMEM;
 121
 122         RB_CLEAR_NODE(&machine->rb_node);
 123         dsos__init(&machine->dsos);
 124
 125         machine__threads_init(machine);
 126
 127         machine->vdso_info = NULL;
 128         machine->env = NULL;
 129
 130         machine->pid = pid;
 131
 132         machine->id_hdr_size = 0;
 133         machine->kptr_restrict_warned = false;
 134         machine->comm_exec = false;
 135         machine->kernel_start = 0;
 136         machine->vmlinux_map = NULL;
 137
 138         machine->root_dir = strdup(root_dir);
 139         if (machine->root_dir == NULL)
 140                 goto out;
 141
 142         if (machine__set_mmap_name(machine))
 143                 goto out;
 144
 145         if (pid != HOST_KERNEL_ID) {
 146                 struct thread *thread = machine__findnew_thread(machine, -1,
 147                                                                 pid);
 148
 149                 if (thread == NULL)
 150                         goto out;
 151
 152                 thread__set_guest_comm(thread, pid);
 153                 thread__put(thread);
 154         }
 155
 156         machine->current_tid = NULL;
 157         err = 0;
 158
 159 out:
 160         if (err) {
 161                 zfree(&machine->kmaps);
 162                 zfree(&machine->root_dir);
 163                 zfree(&machine->mmap_name);
 164         }
 165         return 0;
 166 }
 167
 168 struct machine *machine__new_host(void)
 169 {
 170         struct machine *machine = malloc(sizeof(*machine));
 171
 172         if (machine != NULL) {
 173                 machine__init(machine, "", HOST_KERNEL_ID);
 174
 175                 if (machine__create_kernel_maps(machine) < 0)
 176                         goto out_delete;
 177         }
 178
 179         return machine;
 180 out_delete:
 181         free(machine);
 182         return NULL;
 183 }
 184
 185 struct machine *machine__new_kallsyms(void)
 186 {
 187         struct machine *machine = machine__new_host();
 188         /*
 189          * FIXME:
 190          * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
 191          *    ask for not using the kcore parsing code, once this one is fixed
 192          *    to create a map per module.
 193          */
 194         if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
 195                 machine__delete(machine);
 196                 machine = NULL;
 197         }
 198
 199         return machine;
 200 }
 201
 202 static void dsos__purge(struct dsos *dsos)
 203 {
 204         struct dso *pos, *n;
 205
 206         down_write(&dsos->lock);
 207
 208         list_for_each_entry_safe(pos, n, &dsos->head, node) {
 209                 RB_CLEAR_NODE(&pos->rb_node);
 210                 pos->root = NULL;
 211                 list_del_init(&pos->node);
 212                 dso__put(pos);
 213         }
 214
 215         up_write(&dsos->lock);
 216 }
 217
 218 static void dsos__exit(struct dsos *dsos)
 219 {
 220         dsos__purge(dsos);
 221         exit_rwsem(&dsos->lock);
 222 }
 223
 224 void machine__delete_threads(struct machine *machine)
 225 {
 226         struct rb_node *nd;
 227         int i;
 228
 229         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 230                 struct threads *threads = &machine->threads[i];
 231                 down_write(&threads->lock);
 232                 nd = rb_first_cached(&threads->entries);
 233                 while (nd) {
 234                         struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
 235
 236                         nd = rb_next(nd);
 237                         __machine__remove_thread(machine, trb, trb->thread, false);
 238                 }
 239                 up_write(&threads->lock);
 240         }
 241 }
 242
 243 void machine__exit(struct machine *machine)
 244 {
 245         int i;
 246
 247         if (machine == NULL)
 248                 return;
 249
 250         machine__destroy_kernel_maps(machine);
 251         maps__zput(machine->kmaps);
 252         dsos__exit(&machine->dsos);
 253         machine__exit_vdso(machine);
 254         zfree(&machine->root_dir);
 255         zfree(&machine->mmap_name);
 256         zfree(&machine->current_tid);
 257         zfree(&machine->kallsyms_filename);
 258
 259         machine__delete_threads(machine);
 260         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 261                 struct threads *threads = &machine->threads[i];
 262
 263                 exit_rwsem(&threads->lock);
 264         }
 265 }
 266
 267 void machine__delete(struct machine *machine)
 268 {
 269         if (machine) {
 270                 machine__exit(machine);
 271                 free(machine);
 272         }
 273 }
 274
 275 void machines__init(struct machines *machines)
 276 {
 277         machine__init(&machines->host, "", HOST_KERNEL_ID);
 278         machines->guests = RB_ROOT_CACHED;
 279 }
 280
 281 void machines__exit(struct machines *machines)
 282 {
 283         machine__exit(&machines->host);
 284         /* XXX exit guest */
 285 }
 286
 287 struct machine *machines__add(struct machines *machines, pid_t pid,
 288                               const char *root_dir)
 289 {
 290         struct rb_node **p = &machines->guests.rb_root.rb_node;
 291         struct rb_node *parent = NULL;
 292         struct machine *pos, *machine = malloc(sizeof(*machine));
 293         bool leftmost = true;
 294
 295         if (machine == NULL)
 296                 return NULL;
 297
 298         if (machine__init(machine, root_dir, pid) != 0) {
 299                 free(machine);
 300                 return NULL;
 301         }
 302
 303         while (*p != NULL) {
 304                 parent = *p;
 305                 pos = rb_entry(parent, struct machine, rb_node);
 306                 if (pid < pos->pid)
 307                         p = &(*p)->rb_left;
 308                 else {
 309                         p = &(*p)->rb_right;
 310                         leftmost = false;
 311                 }
 312         }
 313
 314         rb_link_node(&machine->rb_node, parent, p);
 315         rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
 316
 317         machine->machines = machines;
 318
 319         return machine;
 320 }
 321
 322 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 323 {
 324         struct rb_node *nd;
 325
 326         machines->host.comm_exec = comm_exec;
 327
 328         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 329                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 330
 331                 machine->comm_exec = comm_exec;
 332         }
 333 }
 334
 335 struct machine *machines__find(struct machines *machines, pid_t pid)
 336 {
 337         struct rb_node **p = &machines->guests.rb_root.rb_node;
 338         struct rb_node *parent = NULL;
 339         struct machine *machine;
 340         struct machine *default_machine = NULL;
 341
 342         if (pid == HOST_KERNEL_ID)
 343                 return &machines->host;
 344
 345         while (*p != NULL) {
 346                 parent = *p;
 347                 machine = rb_entry(parent, struct machine, rb_node);
 348                 if (pid < machine->pid)
 349                         p = &(*p)->rb_left;
 350                 else if (pid > machine->pid)
 351                         p = &(*p)->rb_right;
 352                 else
 353                         return machine;
 354                 if (!machine->pid)
 355                         default_machine = machine;
 356         }
 357
 358         return default_machine;
 359 }
 360
 361 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 362 {
 363         char path[PATH_MAX];
 364         const char *root_dir = "";
 365         struct machine *machine = machines__find(machines, pid);
 366
 367         if (machine && (machine->pid == pid))
 368                 goto out;
 369
 370         if ((pid != HOST_KERNEL_ID) &&
 371             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 372             (symbol_conf.guestmount)) {
 373                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 374                 if (access(path, R_OK)) {
 375                         static struct strlist *seen;
 376
 377                         if (!seen)
 378                                 seen = strlist__new(NULL, NULL);
 379
 380                         if (!strlist__has_entry(seen, path)) {
 381                                 pr_err("Can't access file %s\n", path);
 382                                 strlist__add(seen, path);
 383                         }
 384                         machine = NULL;
 385                         goto out;
 386                 }
 387                 root_dir = path;
 388         }
 389
 390         machine = machines__add(machines, pid, root_dir);
 391 out:
 392         return machine;
 393 }
 394
 395 struct machine *machines__find_guest(struct machines *machines, pid_t pid)
 396 {
 397         struct machine *machine = machines__find(machines, pid);
 398
 399         if (!machine)
 400                 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
 401         return machine;
 402 }
 403
 404 /*
 405  * A common case for KVM test programs is that the test program acts as the
 406  * hypervisor, creating, running and destroying the virtual machine, and
 407  * providing the guest object code from its own object code. In this case,
 408  * the VM is not running an OS, but only the functions loaded into it by the
 409  * hypervisor test program, and conveniently, loaded at the same virtual
 410  * addresses.
 411  *
 412  * Normally to resolve addresses, MMAP events are needed to map addresses
 413  * back to the object code and debug symbols for that object code.
 414  *
 415  * Currently, there is no way to get such mapping information from guests
 416  * but, in the scenario described above, the guest has the same mappings
 417  * as the hypervisor, so support for that scenario can be achieved.
 418  *
 419  * To support that, copy the host thread's maps to the guest thread's maps.
 420  * Note, we do not discover the guest until we encounter a guest event,
 421  * which works well because it is not until then that we know that the host
 422  * thread's maps have been set up.
 423  *
 424  * This function returns the guest thread. Apart from keeping the data
 425  * structures sane, using a thread belonging to the guest machine, instead
 426  * of the host thread, allows it to have its own comm (refer
 427  * thread__set_guest_comm()).
 428  */
 429 static struct thread *findnew_guest_code(struct machine *machine,
 430                                          struct machine *host_machine,
 431                                          pid_t pid)
 432 {
 433         struct thread *host_thread;
 434         struct thread *thread;
 435         int err;
 436
 437         if (!machine)
 438                 return NULL;
 439
 440         thread = machine__findnew_thread(machine, -1, pid);
 441         if (!thread)
 442                 return NULL;
 443
 444         /* Assume maps are set up if there are any */
 445         if (maps__nr_maps(thread__maps(thread)))
 446                 return thread;
 447
 448         host_thread = machine__find_thread(host_machine, -1, pid);
 449         if (!host_thread)
 450                 goto out_err;
 451
 452         thread__set_guest_comm(thread, pid);
 453
 454         /*
 455          * Guest code can be found in hypervisor process at the same address
 456          * so copy host maps.
 457          */
 458         err = maps__clone(thread, thread__maps(host_thread));
 459         thread__put(host_thread);
 460         if (err)
 461                 goto out_err;
 462
 463         return thread;
 464
 465 out_err:
 466         thread__zput(thread);
 467         return NULL;
 468 }
 469
 470 struct thread *machines__findnew_guest_code(struct machines *machines, pid_t pid)
 471 {
 472         struct machine *host_machine = machines__find(machines, HOST_KERNEL_ID);
 473         struct machine *machine = machines__findnew(machines, pid);
 474
 475         return findnew_guest_code(machine, host_machine, pid);
 476 }
 477
 478 struct thread *machine__findnew_guest_code(struct machine *machine, pid_t pid)
 479 {
 480         struct machines *machines = machine->machines;
 481         struct machine *host_machine;
 482
 483         if (!machines)
 484                 return NULL;
 485
 486         host_machine = machines__find(machines, HOST_KERNEL_ID);
 487
 488         return findnew_guest_code(machine, host_machine, pid);
 489 }
 490
 491 void machines__process_guests(struct machines *machines,
 492                               machine__process_t process, void *data)
 493 {
 494         struct rb_node *nd;
 495
 496         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 497                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 498                 process(pos, data);
 499         }
 500 }
 501
 502 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 503 {
 504         struct rb_node *node;
 505         struct machine *machine;
 506
 507         machines->host.id_hdr_size = id_hdr_size;
 508
 509         for (node = rb_first_cached(&machines->guests); node;
 510              node = rb_next(node)) {
 511                 machine = rb_entry(node, struct machine, rb_node);
 512                 machine->id_hdr_size = id_hdr_size;
 513         }
 514
 515         return;
 516 }
 517
 518 static void machine__update_thread_pid(struct machine *machine,
 519                                        struct thread *th, pid_t pid)
 520 {
 521         struct thread *leader;
 522
 523         if (pid == thread__pid(th) || pid == -1 || thread__pid(th) != -1)
 524                 return;
 525
 526         thread__set_pid(th, pid);
 527
 528         if (thread__pid(th) == thread__tid(th))
 529                 return;
 530
 531         leader = __machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
 532         if (!leader)
 533                 goto out_err;
 534
 535         if (!thread__maps(leader))
 536                 thread__set_maps(leader, maps__new(machine));
 537
 538         if (!thread__maps(leader))
 539                 goto out_err;
 540
 541         if (thread__maps(th) == thread__maps(leader))
 542                 goto out_put;
 543
 544         if (thread__maps(th)) {
 545                 /*
 546                  * Maps are created from MMAP events which provide the pid and
 547                  * tid.  Consequently there never should be any maps on a thread
 548                  * with an unknown pid.  Just print an error if there are.
 549                  */
 550                 if (!maps__empty(thread__maps(th)))
 551                         pr_err("Discarding thread maps for %d:%d\n",
 552                                 thread__pid(th), thread__tid(th));
 553                 maps__put(thread__maps(th));
 554         }
 555
 556         thread__set_maps(th, maps__get(thread__maps(leader)));
 557 out_put:
 558         thread__put(leader);
 559         return;
 560 out_err:
 561         pr_err("Failed to join map groups for %d:%d\n", thread__pid(th), thread__tid(th));
 562         goto out_put;
 563 }
 564
 565 /*
 566  * Front-end cache - TID lookups come in blocks,
 567  * so most of the time we dont have to look up
 568  * the full rbtree:
 569  */
 570 static struct thread*
 571 __threads__get_last_match(struct threads *threads, struct machine *machine,
 572                           int pid, int tid)
 573 {
 574         struct thread *th;
 575
 576         th = threads->last_match;
 577         if (th != NULL) {
 578                 if (thread__tid(th) == tid) {
 579                         machine__update_thread_pid(machine, th, pid);
 580                         return thread__get(th);
 581                 }
 582                 thread__put(threads->last_match);
 583                 threads->last_match = NULL;
 584         }
 585
 586         return NULL;
 587 }
 588
 589 static struct thread*
 590 threads__get_last_match(struct threads *threads, struct machine *machine,
 591                         int pid, int tid)
 592 {
 593         struct thread *th = NULL;
 594
 595         if (perf_singlethreaded)
 596                 th = __threads__get_last_match(threads, machine, pid, tid);
 597
 598         return th;
 599 }
 600
 601 static void
 602 __threads__set_last_match(struct threads *threads, struct thread *th)
 603 {
 604         thread__put(threads->last_match);
 605         threads->last_match = thread__get(th);
 606 }
 607
 608 static void
 609 threads__set_last_match(struct threads *threads, struct thread *th)
 610 {
 611         if (perf_singlethreaded)
 612                 __threads__set_last_match(threads, th);
 613 }
 614
 615 /*
 616  * Caller must eventually drop thread->refcnt returned with a successful
 617  * lookup/new thread inserted.
 618  */
 619 static struct thread *____machine__findnew_thread(struct machine *machine,
 620                                                   struct threads *threads,
 621                                                   pid_t pid, pid_t tid,
 622                                                   bool create)
 623 {
 624         struct rb_node **p = &threads->entries.rb_root.rb_node;
 625         struct rb_node *parent = NULL;
 626         struct thread *th;
 627         struct thread_rb_node *nd;
 628         bool leftmost = true;
 629
 630         th = threads__get_last_match(threads, machine, pid, tid);
 631         if (th)
 632                 return th;
 633
 634         while (*p != NULL) {
 635                 parent = *p;
 636                 th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
 637
 638                 if (thread__tid(th) == tid) {
 639                         threads__set_last_match(threads, th);
 640                         machine__update_thread_pid(machine, th, pid);
 641                         return thread__get(th);
 642                 }
 643
 644                 if (tid < thread__tid(th))
 645                         p = &(*p)->rb_left;
 646                 else {
 647                         p = &(*p)->rb_right;
 648                         leftmost = false;
 649                 }
 650         }
 651
 652         if (!create)
 653                 return NULL;
 654
 655         th = thread__new(pid, tid);
 656         if (th == NULL)
 657                 return NULL;
 658
 659         nd = malloc(sizeof(*nd));
 660         if (nd == NULL) {
 661                 thread__put(th);
 662                 return NULL;
 663         }
 664         nd->thread = th;
 665
 666         rb_link_node(&nd->rb_node, parent, p);
 667         rb_insert_color_cached(&nd->rb_node, &threads->entries, leftmost);
 668         /*
 669          * We have to initialize maps separately after rb tree is updated.
 670          *
 671          * The reason is that we call machine__findnew_thread within
 672          * thread__init_maps to find the thread leader and that would screwed
 673          * the rb tree.
 674          */
 675         if (thread__init_maps(th, machine)) {
 676                 pr_err("Thread init failed thread %d\n", pid);
 677                 rb_erase_cached(&nd->rb_node, &threads->entries);
 678                 RB_CLEAR_NODE(&nd->rb_node);
 679                 free(nd);
 680                 thread__put(th);
 681                 return NULL;
 682         }
 683         /*
 684          * It is now in the rbtree, get a ref
 685          */
 686         threads__set_last_match(threads, th);
 687         ++threads->nr;
 688
 689         return thread__get(th);
 690 }
 691
 692 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 693 {
 694         return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
 695 }
 696
 697 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 698                                        pid_t tid)
 699 {
 700         struct threads *threads = machine__threads(machine, tid);
 701         struct thread *th;
 702
 703         down_write(&threads->lock);
 704         th = __machine__findnew_thread(machine, pid, tid);
 705         up_write(&threads->lock);
 706         return th;
 707 }
 708
 709 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 710                                     pid_t tid)
 711 {
 712         struct threads *threads = machine__threads(machine, tid);
 713         struct thread *th;
 714
 715         down_read(&threads->lock);
 716         th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
 717         up_read(&threads->lock);
 718         return th;
 719 }
 720
 721 /*
 722  * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
 723  * So here a single thread is created for that, but actually there is a separate
 724  * idle task per cpu, so there should be one 'struct thread' per cpu, but there
 725  * is only 1. That causes problems for some tools, requiring workarounds. For
 726  * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
 727  */
 728 struct thread *machine__idle_thread(struct machine *machine)
 729 {
 730         struct thread *thread = machine__findnew_thread(machine, 0, 0);
 731
 732         if (!thread || thread__set_comm(thread, "swapper", 0) ||
 733             thread__set_namespaces(thread, 0, NULL))
 734                 pr_err("problem inserting idle task for machine pid %d\n", machine->pid);
 735
 736         return thread;
 737 }
 738
 739 struct comm *machine__thread_exec_comm(struct machine *machine,
 740                                        struct thread *thread)
 741 {
 742         if (machine->comm_exec)
 743                 return thread__exec_comm(thread);
 744         else
 745                 return thread__comm(thread);
 746 }
 747
 748 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 749                                 struct perf_sample *sample)
 750 {
 751         struct thread *thread = machine__findnew_thread(machine,
 752                                                         event->comm.pid,
 753                                                         event->comm.tid);
 754         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 755         int err = 0;
 756
 757         if (exec)
 758                 machine->comm_exec = true;
 759
 760         if (dump_trace)
 761                 perf_event__fprintf_comm(event, stdout);
 762
 763         if (thread == NULL ||
 764             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 765                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 766                 err = -1;
 767         }
 768
 769         thread__put(thread);
 770
 771         return err;
 772 }
 773
 774 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
 775                                       union perf_event *event,
 776                                       struct perf_sample *sample __maybe_unused)
 777 {
 778         struct thread *thread = machine__findnew_thread(machine,
 779                                                         event->namespaces.pid,
 780                                                         event->namespaces.tid);
 781         int err = 0;
 782
 783         WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
 784                   "\nWARNING: kernel seems to support more namespaces than perf"
 785                   " tool.\nTry updating the perf tool..\n\n");
 786
 787         WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
 788                   "\nWARNING: perf tool seems to support more namespaces than"
 789                   " the kernel.\nTry updating the kernel..\n\n");
 790
 791         if (dump_trace)
 792                 perf_event__fprintf_namespaces(event, stdout);
 793
 794         if (thread == NULL ||
 795             thread__set_namespaces(thread, sample->time, &event->namespaces)) {
 796                 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
 797                 err = -1;
 798         }
 799
 800         thread__put(thread);
 801
 802         return err;
 803 }
 804
 805 int machine__process_cgroup_event(struct machine *machine,
 806                                   union perf_event *event,
 807                                   struct perf_sample *sample __maybe_unused)
 808 {
 809         struct cgroup *cgrp;
 810
 811         if (dump_trace)
 812                 perf_event__fprintf_cgroup(event, stdout);
 813
 814         cgrp = cgroup__findnew(machine->env, event->cgroup.id, event->cgroup.path);
 815         if (cgrp == NULL)
 816                 return -ENOMEM;
 817
 818         return 0;
 819 }
 820
 821 int machine__process_lost_event(struct machine *machine __maybe_unused,
 822                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 823 {
 824         dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
 825                     event->lost.id, event->lost.lost);
 826         return 0;
 827 }
 828
 829 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 830                                         union perf_event *event, struct perf_sample *sample)
 831 {
 832         dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
 833                     sample->id, event->lost_samples.lost);
 834         return 0;
 835 }
 836
 837 static struct dso *machine__findnew_module_dso(struct machine *machine,
 838                                                struct kmod_path *m,
 839                                                const char *filename)
 840 {
 841         struct dso *dso;
 842
 843         down_write(&machine->dsos.lock);
 844
 845         dso = __dsos__find(&machine->dsos, m->name, true);
 846         if (!dso) {
 847                 dso = __dsos__addnew(&machine->dsos, m->name);
 848                 if (dso == NULL)
 849                         goto out_unlock;
 850
 851                 dso__set_module_info(dso, m, machine);
 852                 dso__set_long_name(dso, strdup(filename), true);
 853                 dso->kernel = DSO_SPACE__KERNEL;
 854         }
 855
 856         dso__get(dso);
 857 out_unlock:
 858         up_write(&machine->dsos.lock);
 859         return dso;
 860 }
 861
 862 int machine__process_aux_event(struct machine *machine __maybe_unused,
 863                                union perf_event *event)
 864 {
 865         if (dump_trace)
 866                 perf_event__fprintf_aux(event, stdout);
 867         return 0;
 868 }
 869
 870 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 871                                         union perf_event *event)
 872 {
 873         if (dump_trace)
 874                 perf_event__fprintf_itrace_start(event, stdout);
 875         return 0;
 876 }
 877
 878 int machine__process_aux_output_hw_id_event(struct machine *machine __maybe_unused,
 879                                             union perf_event *event)
 880 {
 881         if (dump_trace)
 882                 perf_event__fprintf_aux_output_hw_id(event, stdout);
 883         return 0;
 884 }
 885
 886 int machine__process_switch_event(struct machine *machine __maybe_unused,
 887                                   union perf_event *event)
 888 {
 889         if (dump_trace)
 890                 perf_event__fprintf_switch(event, stdout);
 891         return 0;
 892 }
 893
 894 static int machine__process_ksymbol_register(struct machine *machine,
 895                                              union perf_event *event,
 896                                              struct perf_sample *sample __maybe_unused)
 897 {
 898         struct symbol *sym;
 899         struct dso *dso;
 900         struct map *map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
 901         bool put_map = false;
 902         int err = 0;
 903
 904         if (!map) {
 905                 dso = dso__new(event->ksymbol.name);
 906
 907                 if (!dso) {
 908                         err = -ENOMEM;
 909                         goto out;
 910                 }
 911                 dso->kernel = DSO_SPACE__KERNEL;
 912                 map = map__new2(0, dso);
 913                 dso__put(dso);
 914                 if (!map) {
 915                         err = -ENOMEM;
 916                         goto out;
 917                 }
 918                 /*
 919                  * The inserted map has a get on it, we need to put to release
 920                  * the reference count here, but do it after all accesses are
 921                  * done.
 922                  */
 923                 put_map = true;
 924                 if (event->ksymbol.ksym_type == PERF_RECORD_KSYMBOL_TYPE_OOL) {
 925                         dso->binary_type = DSO_BINARY_TYPE__OOL;
 926                         dso->data.file_size = event->ksymbol.len;
 927                         dso__set_loaded(dso);
 928                 }
 929
 930                 map__set_start(map, event->ksymbol.addr);
 931                 map__set_end(map, map__start(map) + event->ksymbol.len);
 932                 err = maps__insert(machine__kernel_maps(machine), map);
 933                 if (err) {
 934                         err = -ENOMEM;
 935                         goto out;
 936                 }
 937
 938                 dso__set_loaded(dso);
 939
 940                 if (is_bpf_image(event->ksymbol.name)) {
 941                         dso->binary_type = DSO_BINARY_TYPE__BPF_IMAGE;
 942                         dso__set_long_name(dso, "", false);
 943                 }
 944         } else {
 945                 dso = map__dso(map);
 946         }
 947
 948         sym = symbol__new(map__map_ip(map, map__start(map)),
 949                           event->ksymbol.len,
 950                           0, 0, event->ksymbol.name);
 951         if (!sym) {
 952                 err = -ENOMEM;
 953                 goto out;
 954         }
 955         dso__insert_symbol(dso, sym);
 956 out:
 957         if (put_map)
 958                 map__put(map);
 959         return err;
 960 }
 961
 962 static int machine__process_ksymbol_unregister(struct machine *machine,
 963                                                union perf_event *event,
 964                                                struct perf_sample *sample __maybe_unused)
 965 {
 966         struct symbol *sym;
 967         struct map *map;
 968
 969         map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
 970         if (!map)
 971                 return 0;
 972
 973         if (RC_CHK_ACCESS(map) != RC_CHK_ACCESS(machine->vmlinux_map))
 974                 maps__remove(machine__kernel_maps(machine), map);
 975         else {
 976                 struct dso *dso = map__dso(map);
 977
 978                 sym = dso__find_symbol(dso, map__map_ip(map, map__start(map)));
 979                 if (sym)
 980                         dso__delete_symbol(dso, sym);
 981         }
 982
 983         return 0;
 984 }
 985
 986 int machine__process_ksymbol(struct machine *machine __maybe_unused,
 987                              union perf_event *event,
 988                              struct perf_sample *sample)
 989 {
 990         if (dump_trace)
 991                 perf_event__fprintf_ksymbol(event, stdout);
 992
 993         if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
 994                 return machine__process_ksymbol_unregister(machine, event,
 995                                                            sample);
 996         return machine__process_ksymbol_register(machine, event, sample);
 997 }
 998
 999 int machine__process_text_poke(struct machine *machine, union perf_event *event,
1000                                struct perf_sample *sample __maybe_unused)
1001 {
1002         struct map *map = maps__find(machine__kernel_maps(machine), event->text_poke.addr);
1003         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1004         struct dso *dso = map ? map__dso(map) : NULL;
1005
1006         if (dump_trace)
1007                 perf_event__fprintf_text_poke(event, machine, stdout);
1008
1009         if (!event->text_poke.new_len)
1010                 return 0;
1011
1012         if (cpumode != PERF_RECORD_MISC_KERNEL) {
1013                 pr_debug("%s: unsupported cpumode - ignoring\n", __func__);
1014                 return 0;
1015         }
1016
1017         if (dso) {
1018                 u8 *new_bytes = event->text_poke.bytes + event->text_poke.old_len;
1019                 int ret;
1020
1021                 /*
1022                  * Kernel maps might be changed when loading symbols so loading
1023                  * must be done prior to using kernel maps.
1024                  */
1025                 map__load(map);
1026                 ret = dso__data_write_cache_addr(dso, map, machine,
1027                                                  event->text_poke.addr,
1028                                                  new_bytes,
1029                                                  event->text_poke.new_len);
1030                 if (ret != event->text_poke.new_len)
1031                         pr_debug("Failed to write kernel text poke at %#" PRI_lx64 "\n",
1032                                  event->text_poke.addr);
1033         } else {
1034                 pr_debug("Failed to find kernel text poke address map for %#" PRI_lx64 "\n",
1035                          event->text_poke.addr);
1036         }
1037
1038         return 0;
1039 }
1040
1041 static struct map *machine__addnew_module_map(struct machine *machine, u64 start,
1042                                               const char *filename)
1043 {
1044         struct map *map = NULL;
1045         struct kmod_path m;
1046         struct dso *dso;
1047         int err;
1048
1049         if (kmod_path__parse_name(&m, filename))
1050                 return NULL;
1051
1052         dso = machine__findnew_module_dso(machine, &m, filename);
1053         if (dso == NULL)
1054                 goto out;
1055
1056         map = map__new2(start, dso);
1057         if (map == NULL)
1058                 goto out;
1059
1060         err = maps__insert(machine__kernel_maps(machine), map);
1061         /* If maps__insert failed, return NULL. */
1062         if (err) {
1063                 map__put(map);
1064                 map = NULL;
1065         }
1066 out:
1067         /* put the dso here, corresponding to  machine__findnew_module_dso */
1068         dso__put(dso);
1069         zfree(&m.name);
1070         return map;
1071 }
1072
1073 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
1074 {
1075         struct rb_node *nd;
1076         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
1077
1078         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
1079                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
1080                 ret += __dsos__fprintf(&pos->dsos.head, fp);
1081         }
1082
1083         return ret;
1084 }
1085
1086 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
1087                                      bool (skip)(struct dso *dso, int parm), int parm)
1088 {
1089         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
1090 }
1091
1092 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
1093                                      bool (skip)(struct dso *dso, int parm), int parm)
1094 {
1095         struct rb_node *nd;
1096         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
1097
1098         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
1099                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
1100                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
1101         }
1102         return ret;
1103 }
1104
1105 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
1106 {
1107         int i;
1108         size_t printed = 0;
1109         struct dso *kdso = machine__kernel_dso(machine);
1110
1111         if (kdso->has_build_id) {
1112                 char filename[PATH_MAX];
1113                 if (dso__build_id_filename(kdso, filename, sizeof(filename),
1114                                            false))
1115                         printed += fprintf(fp, "[0] %s\n", filename);
1116         }
1117
1118         for (i = 0; i < vmlinux_path__nr_entries; ++i)
1119                 printed += fprintf(fp, "[%d] %s\n",
1120                                    i + kdso->has_build_id, vmlinux_path[i]);
1121
1122         return printed;
1123 }
1124
1125 size_t machine__fprintf(struct machine *machine, FILE *fp)
1126 {
1127         struct rb_node *nd;
1128         size_t ret;
1129         int i;
1130
1131         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
1132                 struct threads *threads = &machine->threads[i];
1133
1134                 down_read(&threads->lock);
1135
1136                 ret = fprintf(fp, "Threads: %u\n", threads->nr);
1137
1138                 for (nd = rb_first_cached(&threads->entries); nd;
1139                      nd = rb_next(nd)) {
1140                         struct thread *pos = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
1141
1142                         ret += thread__fprintf(pos, fp);
1143                 }
1144
1145                 up_read(&threads->lock);
1146         }
1147         return ret;
1148 }
1149
1150 static struct dso *machine__get_kernel(struct machine *machine)
1151 {
1152         const char *vmlinux_name = machine->mmap_name;
1153         struct dso *kernel;
1154
1155         if (machine__is_host(machine)) {
1156                 if (symbol_conf.vmlinux_name)
1157                         vmlinux_name = symbol_conf.vmlinux_name;
1158
1159                 kernel = machine__findnew_kernel(machine, vmlinux_name,
1160                                                  "[kernel]", DSO_SPACE__KERNEL);
1161         } else {
1162                 if (symbol_conf.default_guest_vmlinux_name)
1163                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
1164
1165                 kernel = machine__findnew_kernel(machine, vmlinux_name,
1166                                                  "[guest.kernel]",
1167                                                  DSO_SPACE__KERNEL_GUEST);
1168         }
1169
1170         if (kernel != NULL && (!kernel->has_build_id))
1171                 dso__read_running_kernel_build_id(kernel, machine);
1172
1173         return kernel;
1174 }
1175
1176 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
1177                                     size_t bufsz)
1178 {
1179         if (machine__is_default_guest(machine))
1180                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
1181         else
1182                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
1183 }
1184
1185 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
1186
1187 /* Figure out the start address of kernel map from /proc/kallsyms.
1188  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
1189  * symbol_name if it's not that important.
1190  */
1191 static int machine__get_running_kernel_start(struct machine *machine,
1192                                              const char **symbol_name,
1193                                              u64 *start, u64 *end)
1194 {
1195         char filename[PATH_MAX];
1196         int i, err = -1;
1197         const char *name;
1198         u64 addr = 0;
1199
1200         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
1201
1202         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
1203                 return 0;
1204
1205         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
1206                 err = kallsyms__get_function_start(filename, name, &addr);
1207                 if (!err)
1208                         break;
1209         }
1210
1211         if (err)
1212                 return -1;
1213
1214         if (symbol_name)
1215                 *symbol_name = name;
1216
1217         *start = addr;
1218
1219         err = kallsyms__get_symbol_start(filename, "_edata", &addr);
1220         if (err)
1221                 err = kallsyms__get_function_start(filename, "_etext", &addr);
1222         if (!err)
1223                 *end = addr;
1224
1225         return 0;
1226 }
1227
1228 int machine__create_extra_kernel_map(struct machine *machine,
1229                                      struct dso *kernel,
1230                                      struct extra_kernel_map *xm)
1231 {
1232         struct kmap *kmap;
1233         struct map *map;
1234         int err;
1235
1236         map = map__new2(xm->start, kernel);
1237         if (!map)
1238                 return -ENOMEM;
1239
1240         map__set_end(map, xm->end);
1241         map__set_pgoff(map, xm->pgoff);
1242
1243         kmap = map__kmap(map);
1244
1245         strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1246
1247         err = maps__insert(machine__kernel_maps(machine), map);
1248
1249         if (!err) {
1250                 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1251                         kmap->name, map__start(map), map__end(map));
1252         }
1253
1254         map__put(map);
1255
1256         return err;
1257 }
1258
1259 static u64 find_entry_trampoline(struct dso *dso)
1260 {
1261         /* Duplicates are removed so lookup all aliases */
1262         const char *syms[] = {
1263                 "_entry_trampoline",
1264                 "__entry_trampoline_start",
1265                 "entry_SYSCALL_64_trampoline",
1266         };
1267         struct symbol *sym = dso__first_symbol(dso);
1268         unsigned int i;
1269
1270         for (; sym; sym = dso__next_symbol(sym)) {
1271                 if (sym->binding != STB_GLOBAL)
1272                         continue;
1273                 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1274                         if (!strcmp(sym->name, syms[i]))
1275                                 return sym->start;
1276                 }
1277         }
1278
1279         return 0;
1280 }
1281
1282 /*
1283  * These values can be used for kernels that do not have symbols for the entry
1284  * trampolines in kallsyms.
1285  */
1286 #define X86_64_CPU_ENTRY_AREA_PER_CPU   0xfffffe0000000000ULL
1287 #define X86_64_CPU_ENTRY_AREA_SIZE      0x2c000
1288 #define X86_64_ENTRY_TRAMPOLINE         0x6000
1289
1290 /* Map x86_64 PTI entry trampolines */
1291 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1292                                           struct dso *kernel)
1293 {
1294         struct maps *kmaps = machine__kernel_maps(machine);
1295         int nr_cpus_avail, cpu;
1296         bool found = false;
1297         struct map_rb_node *rb_node;
1298         u64 pgoff;
1299
1300         /*
1301          * In the vmlinux case, pgoff is a virtual address which must now be
1302          * mapped to a vmlinux offset.
1303          */
1304         maps__for_each_entry(kmaps, rb_node) {
1305                 struct map *dest_map, *map = rb_node->map;
1306                 struct kmap *kmap = __map__kmap(map);
1307
1308                 if (!kmap || !is_entry_trampoline(kmap->name))
1309                         continue;
1310
1311                 dest_map = maps__find(kmaps, map__pgoff(map));
1312                 if (dest_map != map)
1313                         map__set_pgoff(map, map__map_ip(dest_map, map__pgoff(map)));
1314                 found = true;
1315         }
1316         if (found || machine->trampolines_mapped)
1317                 return 0;
1318
1319         pgoff = find_entry_trampoline(kernel);
1320         if (!pgoff)
1321                 return 0;
1322
1323         nr_cpus_avail = machine__nr_cpus_avail(machine);
1324
1325         /* Add a 1 page map for each CPU's entry trampoline */
1326         for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1327                 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1328                          cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1329                          X86_64_ENTRY_TRAMPOLINE;
1330                 struct extra_kernel_map xm = {
1331                         .start = va,
1332                         .end   = va + page_size,
1333                         .pgoff = pgoff,
1334                 };
1335
1336                 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1337
1338                 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1339                         return -1;
1340         }
1341
1342         machine->trampolines_mapped = nr_cpus_avail;
1343
1344         return 0;
1345 }
1346
1347 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1348                                              struct dso *kernel __maybe_unused)
1349 {
1350         return 0;
1351 }
1352
1353 static int
1354 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1355 {
1356         /* In case of renewal the kernel map, destroy previous one */
1357         machine__destroy_kernel_maps(machine);
1358
1359         map__put(machine->vmlinux_map);
1360         machine->vmlinux_map = map__new2(0, kernel);
1361         if (machine->vmlinux_map == NULL)
1362                 return -ENOMEM;
1363
1364         map__set_map_ip(machine->vmlinux_map, identity__map_ip);
1365         map__set_unmap_ip(machine->vmlinux_map, identity__map_ip);
1366         return maps__insert(machine__kernel_maps(machine), machine->vmlinux_map);
1367 }
1368
1369 void machine__destroy_kernel_maps(struct machine *machine)
1370 {
1371         struct kmap *kmap;
1372         struct map *map = machine__kernel_map(machine);
1373
1374         if (map == NULL)
1375                 return;
1376
1377         kmap = map__kmap(map);
1378         maps__remove(machine__kernel_maps(machine), map);
1379         if (kmap && kmap->ref_reloc_sym) {
1380                 zfree((char **)&kmap->ref_reloc_sym->name);
1381                 zfree(&kmap->ref_reloc_sym);
1382         }
1383
1384         map__zput(machine->vmlinux_map);
1385 }
1386
1387 int machines__create_guest_kernel_maps(struct machines *machines)
1388 {
1389         int ret = 0;
1390         struct dirent **namelist = NULL;
1391         int i, items = 0;
1392         char path[PATH_MAX];
1393         pid_t pid;
1394         char *endp;
1395
1396         if (symbol_conf.default_guest_vmlinux_name ||
1397             symbol_conf.default_guest_modules ||
1398             symbol_conf.default_guest_kallsyms) {
1399                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1400         }
1401
1402         if (symbol_conf.guestmount) {
1403                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1404                 if (items <= 0)
1405                         return -ENOENT;
1406                 for (i = 0; i < items; i++) {
1407                         if (!isdigit(namelist[i]->d_name[0])) {
1408                                 /* Filter out . and .. */
1409                                 continue;
1410                         }
1411                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1412                         if ((*endp != '\0') ||
1413                             (endp == namelist[i]->d_name) ||
1414                             (errno == ERANGE)) {
1415                                 pr_debug("invalid directory (%s). Skipping.\n",
1416                                          namelist[i]->d_name);
1417                                 continue;
1418                         }
1419                         sprintf(path, "%s/%s/proc/kallsyms",
1420                                 symbol_conf.guestmount,
1421                                 namelist[i]->d_name);
1422                         ret = access(path, R_OK);
1423                         if (ret) {
1424                                 pr_debug("Can't access file %s\n", path);
1425                                 goto failure;
1426                         }
1427                         machines__create_kernel_maps(machines, pid);
1428                 }
1429 failure:
1430                 free(namelist);
1431         }
1432
1433         return ret;
1434 }
1435
1436 void machines__destroy_kernel_maps(struct machines *machines)
1437 {
1438         struct rb_node *next = rb_first_cached(&machines->guests);
1439
1440         machine__destroy_kernel_maps(&machines->host);
1441
1442         while (next) {
1443                 struct machine *pos = rb_entry(next, struct machine, rb_node);
1444
1445                 next = rb_next(&pos->rb_node);
1446                 rb_erase_cached(&pos->rb_node, &machines->guests);
1447                 machine__delete(pos);
1448         }
1449 }
1450
1451 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1452 {
1453         struct machine *machine = machines__findnew(machines, pid);
1454
1455         if (machine == NULL)
1456                 return -1;
1457
1458         return machine__create_kernel_maps(machine);
1459 }
1460
1461 int machine__load_kallsyms(struct machine *machine, const char *filename)
1462 {
1463         struct map *map = machine__kernel_map(machine);
1464         struct dso *dso = map__dso(map);
1465         int ret = __dso__load_kallsyms(dso, filename, map, true);
1466
1467         if (ret > 0) {
1468                 dso__set_loaded(dso);
1469                 /*
1470                  * Since /proc/kallsyms will have multiple sessions for the
1471                  * kernel, with modules between them, fixup the end of all
1472                  * sections.
1473                  */
1474                 maps__fixup_end(machine__kernel_maps(machine));
1475         }
1476
1477         return ret;
1478 }
1479
1480 int machine__load_vmlinux_path(struct machine *machine)
1481 {
1482         struct map *map = machine__kernel_map(machine);
1483         struct dso *dso = map__dso(map);
1484         int ret = dso__load_vmlinux_path(dso, map);
1485
1486         if (ret > 0)
1487                 dso__set_loaded(dso);
1488
1489         return ret;
1490 }
1491
1492 static char *get_kernel_version(const char *root_dir)
1493 {
1494         char version[PATH_MAX];
1495         FILE *file;
1496         char *name, *tmp;
1497         const char *prefix = "Linux version ";
1498
1499         sprintf(version, "%s/proc/version", root_dir);
1500         file = fopen(version, "r");
1501         if (!file)
1502                 return NULL;
1503
1504         tmp = fgets(version, sizeof(version), file);
1505         fclose(file);
1506         if (!tmp)
1507                 return NULL;
1508
1509         name = strstr(version, prefix);
1510         if (!name)
1511                 return NULL;
1512         name += strlen(prefix);
1513         tmp = strchr(name, ' ');
1514         if (tmp)
1515                 *tmp = '\0';
1516
1517         return strdup(name);
1518 }
1519
1520 static bool is_kmod_dso(struct dso *dso)
1521 {
1522         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1523                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1524 }
1525
1526 static int maps__set_module_path(struct maps *maps, const char *path, struct kmod_path *m)
1527 {
1528         char *long_name;
1529         struct dso *dso;
1530         struct map *map = maps__find_by_name(maps, m->name);
1531
1532         if (map == NULL)
1533                 return 0;
1534
1535         long_name = strdup(path);
1536         if (long_name == NULL)
1537                 return -ENOMEM;
1538
1539         dso = map__dso(map);
1540         dso__set_long_name(dso, long_name, true);
1541         dso__kernel_module_get_build_id(dso, "");
1542
1543         /*
1544          * Full name could reveal us kmod compression, so
1545          * we need to update the symtab_type if needed.
1546          */
1547         if (m->comp && is_kmod_dso(dso)) {
1548                 dso->symtab_type++;
1549                 dso->comp = m->comp;
1550         }
1551
1552         return 0;
1553 }
1554
1555 static int maps__set_modules_path_dir(struct maps *maps, const char *dir_name, int depth)
1556 {
1557         struct dirent *dent;
1558         DIR *dir = opendir(dir_name);
1559         int ret = 0;
1560
1561         if (!dir) {
1562                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1563                 return -1;
1564         }
1565
1566         while ((dent = readdir(dir)) != NULL) {
1567                 char path[PATH_MAX];
1568                 struct stat st;
1569
1570                 /*sshfs might return bad dent->d_type, so we have to stat*/
1571                 path__join(path, sizeof(path), dir_name, dent->d_name);
1572                 if (stat(path, &st))
1573                         continue;
1574
1575                 if (S_ISDIR(st.st_mode)) {
1576                         if (!strcmp(dent->d_name, ".") ||
1577                             !strcmp(dent->d_name, ".."))
1578                                 continue;
1579
1580                         /* Do not follow top-level source and build symlinks */
1581                         if (depth == 0) {
1582                                 if (!strcmp(dent->d_name, "source") ||
1583                                     !strcmp(dent->d_name, "build"))
1584                                         continue;
1585                         }
1586
1587                         ret = maps__set_modules_path_dir(maps, path, depth + 1);
1588                         if (ret < 0)
1589                                 goto out;
1590                 } else {
1591                         struct kmod_path m;
1592
1593                         ret = kmod_path__parse_name(&m, dent->d_name);
1594                         if (ret)
1595                                 goto out;
1596
1597                         if (m.kmod)
1598                                 ret = maps__set_module_path(maps, path, &m);
1599
1600                         zfree(&m.name);
1601
1602                         if (ret)
1603                                 goto out;
1604                 }
1605         }
1606
1607 out:
1608         closedir(dir);
1609         return ret;
1610 }
1611
1612 static int machine__set_modules_path(struct machine *machine)
1613 {
1614         char *version;
1615         char modules_path[PATH_MAX];
1616
1617         version = get_kernel_version(machine->root_dir);
1618         if (!version)
1619                 return -1;
1620
1621         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1622                  machine->root_dir, version);
1623         free(version);
1624
1625         return maps__set_modules_path_dir(machine__kernel_maps(machine), modules_path, 0);
1626 }
1627 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1628                                 u64 *size __maybe_unused,
1629                                 const char *name __maybe_unused)
1630 {
1631         return 0;
1632 }
1633
1634 static int machine__create_module(void *arg, const char *name, u64 start,
1635                                   u64 size)
1636 {
1637         struct machine *machine = arg;
1638         struct map *map;
1639
1640         if (arch__fix_module_text_start(&start, &size, name) < 0)
1641                 return -1;
1642
1643         map = machine__addnew_module_map(machine, start, name);
1644         if (map == NULL)
1645                 return -1;
1646         map__set_end(map, start + size);
1647
1648         dso__kernel_module_get_build_id(map__dso(map), machine->root_dir);
1649         map__put(map);
1650         return 0;
1651 }
1652
1653 static int machine__create_modules(struct machine *machine)
1654 {
1655         const char *modules;
1656         char path[PATH_MAX];
1657
1658         if (machine__is_default_guest(machine)) {
1659                 modules = symbol_conf.default_guest_modules;
1660         } else {
1661                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1662                 modules = path;
1663         }
1664
1665         if (symbol__restricted_filename(modules, "/proc/modules"))
1666                 return -1;
1667
1668         if (modules__parse(modules, machine, machine__create_module))
1669                 return -1;
1670
1671         if (!machine__set_modules_path(machine))
1672                 return 0;
1673
1674         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1675
1676         return 0;
1677 }
1678
1679 static void machine__set_kernel_mmap(struct machine *machine,
1680                                      u64 start, u64 end)
1681 {
1682         map__set_start(machine->vmlinux_map, start);
1683         map__set_end(machine->vmlinux_map, end);
1684         /*
1685          * Be a bit paranoid here, some perf.data file came with
1686          * a zero sized synthesized MMAP event for the kernel.
1687          */
1688         if (start == 0 && end == 0)
1689                 map__set_end(machine->vmlinux_map, ~0ULL);
1690 }
1691
1692 static int machine__update_kernel_mmap(struct machine *machine,
1693                                      u64 start, u64 end)
1694 {
1695         struct map *orig, *updated;
1696         int err;
1697
1698         orig = machine->vmlinux_map;
1699         updated = map__get(orig);
1700
1701         machine->vmlinux_map = updated;
1702         machine__set_kernel_mmap(machine, start, end);
1703         maps__remove(machine__kernel_maps(machine), orig);
1704         err = maps__insert(machine__kernel_maps(machine), updated);
1705         map__put(orig);
1706
1707         return err;
1708 }
1709
1710 int machine__create_kernel_maps(struct machine *machine)
1711 {
1712         struct dso *kernel = machine__get_kernel(machine);
1713         const char *name = NULL;
1714         u64 start = 0, end = ~0ULL;
1715         int ret;
1716
1717         if (kernel == NULL)
1718                 return -1;
1719
1720         ret = __machine__create_kernel_maps(machine, kernel);
1721         if (ret < 0)
1722                 goto out_put;
1723
1724         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1725                 if (machine__is_host(machine))
1726                         pr_debug("Problems creating module maps, "
1727                                  "continuing anyway...\n");
1728                 else
1729                         pr_debug("Problems creating module maps for guest %d, "
1730                                  "continuing anyway...\n", machine->pid);
1731         }
1732
1733         if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1734                 if (name &&
1735                     map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1736                         machine__destroy_kernel_maps(machine);
1737                         ret = -1;
1738                         goto out_put;
1739                 }
1740
1741                 /*
1742                  * we have a real start address now, so re-order the kmaps
1743                  * assume it's the last in the kmaps
1744                  */
1745                 ret = machine__update_kernel_mmap(machine, start, end);
1746                 if (ret < 0)
1747                         goto out_put;
1748         }
1749
1750         if (machine__create_extra_kernel_maps(machine, kernel))
1751                 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1752
1753         if (end == ~0ULL) {
1754                 /* update end address of the kernel map using adjacent module address */
1755                 struct map_rb_node *rb_node = maps__find_node(machine__kernel_maps(machine),
1756                                                         machine__kernel_map(machine));
1757                 struct map_rb_node *next = map_rb_node__next(rb_node);
1758
1759                 if (next)
1760                         machine__set_kernel_mmap(machine, start, map__start(next->map));
1761         }
1762
1763 out_put:
1764         dso__put(kernel);
1765         return ret;
1766 }
1767
1768 static bool machine__uses_kcore(struct machine *machine)
1769 {
1770         struct dso *dso;
1771
1772         list_for_each_entry(dso, &machine->dsos.head, node) {
1773                 if (dso__is_kcore(dso))
1774                         return true;
1775         }
1776
1777         return false;
1778 }
1779
1780 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1781                                              struct extra_kernel_map *xm)
1782 {
1783         return machine__is(machine, "x86_64") &&
1784                is_entry_trampoline(xm->name);
1785 }
1786
1787 static int machine__process_extra_kernel_map(struct machine *machine,
1788                                              struct extra_kernel_map *xm)
1789 {
1790         struct dso *kernel = machine__kernel_dso(machine);
1791
1792         if (kernel == NULL)
1793                 return -1;
1794
1795         return machine__create_extra_kernel_map(machine, kernel, xm);
1796 }
1797
1798 static int machine__process_kernel_mmap_event(struct machine *machine,
1799                                               struct extra_kernel_map *xm,
1800                                               struct build_id *bid)
1801 {
1802         enum dso_space_type dso_space;
1803         bool is_kernel_mmap;
1804         const char *mmap_name = machine->mmap_name;
1805
1806         /* If we have maps from kcore then we do not need or want any others */
1807         if (machine__uses_kcore(machine))
1808                 return 0;
1809
1810         if (machine__is_host(machine))
1811                 dso_space = DSO_SPACE__KERNEL;
1812         else
1813                 dso_space = DSO_SPACE__KERNEL_GUEST;
1814
1815         is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1816         if (!is_kernel_mmap && !machine__is_host(machine)) {
1817                 /*
1818                  * If the event was recorded inside the guest and injected into
1819                  * the host perf.data file, then it will match a host mmap_name,
1820                  * so try that - see machine__set_mmap_name().
1821                  */
1822                 mmap_name = "[kernel.kallsyms]";
1823                 is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1824         }
1825         if (xm->name[0] == '/' ||
1826             (!is_kernel_mmap && xm->name[0] == '[')) {
1827                 struct map *map = machine__addnew_module_map(machine, xm->start, xm->name);
1828
1829                 if (map == NULL)
1830                         goto out_problem;
1831
1832                 map__set_end(map, map__start(map) + xm->end - xm->start);
1833
1834                 if (build_id__is_defined(bid))
1835                         dso__set_build_id(map__dso(map), bid);
1836
1837                 map__put(map);
1838         } else if (is_kernel_mmap) {
1839                 const char *symbol_name = xm->name + strlen(mmap_name);
1840                 /*
1841                  * Should be there already, from the build-id table in
1842                  * the header.
1843                  */
1844                 struct dso *kernel = NULL;
1845                 struct dso *dso;
1846
1847                 down_read(&machine->dsos.lock);
1848
1849                 list_for_each_entry(dso, &machine->dsos.head, node) {
1850
1851                         /*
1852                          * The cpumode passed to is_kernel_module is not the
1853                          * cpumode of *this* event. If we insist on passing
1854                          * correct cpumode to is_kernel_module, we should
1855                          * record the cpumode when we adding this dso to the
1856                          * linked list.
1857                          *
1858                          * However we don't really need passing correct
1859                          * cpumode.  We know the correct cpumode must be kernel
1860                          * mode (if not, we should not link it onto kernel_dsos
1861                          * list).
1862                          *
1863                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1864                          * is_kernel_module() treats it as a kernel cpumode.
1865                          */
1866
1867                         if (!dso->kernel ||
1868                             is_kernel_module(dso->long_name,
1869                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1870                                 continue;
1871
1872
1873                         kernel = dso__get(dso);
1874                         break;
1875                 }
1876
1877                 up_read(&machine->dsos.lock);
1878
1879                 if (kernel == NULL)
1880                         kernel = machine__findnew_dso(machine, machine->mmap_name);
1881                 if (kernel == NULL)
1882                         goto out_problem;
1883
1884                 kernel->kernel = dso_space;
1885                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1886                         dso__put(kernel);
1887                         goto out_problem;
1888                 }
1889
1890                 if (strstr(kernel->long_name, "vmlinux"))
1891                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1892
1893                 if (machine__update_kernel_mmap(machine, xm->start, xm->end) < 0) {
1894                         dso__put(kernel);
1895                         goto out_problem;
1896                 }
1897
1898                 if (build_id__is_defined(bid))
1899                         dso__set_build_id(kernel, bid);
1900
1901                 /*
1902                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1903                  * symbol. Effectively having zero here means that at record
1904                  * time /proc/sys/kernel/kptr_restrict was non zero.
1905                  */
1906                 if (xm->pgoff != 0) {
1907                         map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1908                                                         symbol_name,
1909                                                         xm->pgoff);
1910                 }
1911
1912                 if (machine__is_default_guest(machine)) {
1913                         /*
1914                          * preload dso of guest kernel and modules
1915                          */
1916                         dso__load(kernel, machine__kernel_map(machine));
1917                 }
1918                 dso__put(kernel);
1919         } else if (perf_event__is_extra_kernel_mmap(machine, xm)) {
1920                 return machine__process_extra_kernel_map(machine, xm);
1921         }
1922         return 0;
1923 out_problem:
1924         return -1;
1925 }
1926
1927 int machine__process_mmap2_event(struct machine *machine,
1928                                  union perf_event *event,
1929                                  struct perf_sample *sample)
1930 {
1931         struct thread *thread;
1932         struct map *map;
1933         struct dso_id dso_id = {
1934                 .maj = event->mmap2.maj,
1935                 .min = event->mmap2.min,
1936                 .ino = event->mmap2.ino,
1937                 .ino_generation = event->mmap2.ino_generation,
1938         };
1939         struct build_id __bid, *bid = NULL;
1940         int ret = 0;
1941
1942         if (dump_trace)
1943                 perf_event__fprintf_mmap2(event, stdout);
1944
1945         if (event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID) {
1946                 bid = &__bid;
1947                 build_id__init(bid, event->mmap2.build_id, event->mmap2.build_id_size);
1948         }
1949
1950         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1951             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1952                 struct extra_kernel_map xm = {
1953                         .start = event->mmap2.start,
1954                         .end   = event->mmap2.start + event->mmap2.len,
1955                         .pgoff = event->mmap2.pgoff,
1956                 };
1957
1958                 strlcpy(xm.name, event->mmap2.filename, KMAP_NAME_LEN);
1959                 ret = machine__process_kernel_mmap_event(machine, &xm, bid);
1960                 if (ret < 0)
1961                         goto out_problem;
1962                 return 0;
1963         }
1964
1965         thread = machine__findnew_thread(machine, event->mmap2.pid,
1966                                         event->mmap2.tid);
1967         if (thread == NULL)
1968                 goto out_problem;
1969
1970         map = map__new(machine, event->mmap2.start,
1971                         event->mmap2.len, event->mmap2.pgoff,
1972                         &dso_id, event->mmap2.prot,
1973                         event->mmap2.flags, bid,
1974                         event->mmap2.filename, thread);
1975
1976         if (map == NULL)
1977                 goto out_problem_map;
1978
1979         ret = thread__insert_map(thread, map);
1980         if (ret)
1981                 goto out_problem_insert;
1982
1983         thread__put(thread);
1984         map__put(map);
1985         return 0;
1986
1987 out_problem_insert:
1988         map__put(map);
1989 out_problem_map:
1990         thread__put(thread);
1991 out_problem:
1992         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1993         return 0;
1994 }
1995
1996 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1997                                 struct perf_sample *sample)
1998 {
1999         struct thread *thread;
2000         struct map *map;
2001         u32 prot = 0;
2002         int ret = 0;
2003
2004         if (dump_trace)
2005                 perf_event__fprintf_mmap(event, stdout);
2006
2007         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
2008             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
2009                 struct extra_kernel_map xm = {
2010                         .start = event->mmap.start,
2011                         .end   = event->mmap.start + event->mmap.len,
2012                         .pgoff = event->mmap.pgoff,
2013                 };
2014
2015                 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
2016                 ret = machine__process_kernel_mmap_event(machine, &xm, NULL);
2017                 if (ret < 0)
2018                         goto out_problem;
2019                 return 0;
2020         }
2021
2022         thread = machine__findnew_thread(machine, event->mmap.pid,
2023                                          event->mmap.tid);
2024         if (thread == NULL)
2025                 goto out_problem;
2026
2027         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
2028                 prot = PROT_EXEC;
2029
2030         map = map__new(machine, event->mmap.start,
2031                         event->mmap.len, event->mmap.pgoff,
2032                         NULL, prot, 0, NULL, event->mmap.filename, thread);
2033
2034         if (map == NULL)
2035                 goto out_problem_map;
2036
2037         ret = thread__insert_map(thread, map);
2038         if (ret)
2039                 goto out_problem_insert;
2040
2041         thread__put(thread);
2042         map__put(map);
2043         return 0;
2044
2045 out_problem_insert:
2046         map__put(map);
2047 out_problem_map:
2048         thread__put(thread);
2049 out_problem:
2050         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
2051         return 0;
2052 }
2053
2054 static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
2055                                      struct thread *th, bool lock)
2056 {
2057         struct threads *threads = machine__threads(machine, thread__tid(th));
2058
2059         if (!nd)
2060                 nd = thread_rb_node__find(th, &threads->entries.rb_root);
2061
2062         if (threads->last_match && RC_CHK_ACCESS(threads->last_match) == RC_CHK_ACCESS(th))
2063                 threads__set_last_match(threads, NULL);
2064
2065         if (lock)
2066                 down_write(&threads->lock);
2067
2068         BUG_ON(refcount_read(thread__refcnt(th)) == 0);
2069
2070         thread__put(nd->thread);
2071         rb_erase_cached(&nd->rb_node, &threads->entries);
2072         RB_CLEAR_NODE(&nd->rb_node);
2073         --threads->nr;
2074
2075         free(nd);
2076
2077         if (lock)
2078                 up_write(&threads->lock);
2079 }
2080
2081 void machine__remove_thread(struct machine *machine, struct thread *th)
2082 {
2083         return __machine__remove_thread(machine, NULL, th, true);
2084 }
2085
2086 int machine__process_fork_event(struct machine *machine, union perf_event *event,
2087                                 struct perf_sample *sample)
2088 {
2089         struct thread *thread = machine__find_thread(machine,
2090                                                      event->fork.pid,
2091                                                      event->fork.tid);
2092         struct thread *parent = machine__findnew_thread(machine,
2093                                                         event->fork.ppid,
2094                                                         event->fork.ptid);
2095         bool do_maps_clone = true;
2096         int err = 0;
2097
2098         if (dump_trace)
2099                 perf_event__fprintf_task(event, stdout);
2100
2101         /*
2102          * There may be an existing thread that is not actually the parent,
2103          * either because we are processing events out of order, or because the
2104          * (fork) event that would have removed the thread was lost. Assume the
2105          * latter case and continue on as best we can.
2106          */
2107         if (thread__pid(parent) != (pid_t)event->fork.ppid) {
2108                 dump_printf("removing erroneous parent thread %d/%d\n",
2109                             thread__pid(parent), thread__tid(parent));
2110                 machine__remove_thread(machine, parent);
2111                 thread__put(parent);
2112                 parent = machine__findnew_thread(machine, event->fork.ppid,
2113                                                  event->fork.ptid);
2114         }
2115
2116         /* if a thread currently exists for the thread id remove it */
2117         if (thread != NULL) {
2118                 machine__remove_thread(machine, thread);
2119                 thread__put(thread);
2120         }
2121
2122         thread = machine__findnew_thread(machine, event->fork.pid,
2123                                          event->fork.tid);
2124         /*
2125          * When synthesizing FORK events, we are trying to create thread
2126          * objects for the already running tasks on the machine.
2127          *
2128          * Normally, for a kernel FORK event, we want to clone the parent's
2129          * maps because that is what the kernel just did.
2130          *
2131          * But when synthesizing, this should not be done.  If we do, we end up
2132          * with overlapping maps as we process the synthesized MMAP2 events that
2133          * get delivered shortly thereafter.
2134          *
2135          * Use the FORK event misc flags in an internal way to signal this
2136          * situation, so we can elide the map clone when appropriate.
2137          */
2138         if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
2139                 do_maps_clone = false;
2140
2141         if (thread == NULL || parent == NULL ||
2142             thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
2143                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
2144                 err = -1;
2145         }
2146         thread__put(thread);
2147         thread__put(parent);
2148
2149         return err;
2150 }
2151
2152 int machine__process_exit_event(struct machine *machine, union perf_event *event,
2153                                 struct perf_sample *sample __maybe_unused)
2154 {
2155         struct thread *thread = machine__find_thread(machine,
2156                                                      event->fork.pid,
2157                                                      event->fork.tid);
2158
2159         if (dump_trace)
2160                 perf_event__fprintf_task(event, stdout);
2161
2162         if (thread != NULL)
2163                 thread__put(thread);
2164
2165         return 0;
2166 }
2167
2168 int machine__process_event(struct machine *machine, union perf_event *event,
2169                            struct perf_sample *sample)
2170 {
2171         int ret;
2172
2173         switch (event->header.type) {
2174         case PERF_RECORD_COMM:
2175                 ret = machine__process_comm_event(machine, event, sample); break;
2176         case PERF_RECORD_MMAP:
2177                 ret = machine__process_mmap_event(machine, event, sample); break;
2178         case PERF_RECORD_NAMESPACES:
2179                 ret = machine__process_namespaces_event(machine, event, sample); break;
2180         case PERF_RECORD_CGROUP:
2181                 ret = machine__process_cgroup_event(machine, event, sample); break;
2182         case PERF_RECORD_MMAP2:
2183                 ret = machine__process_mmap2_event(machine, event, sample); break;
2184         case PERF_RECORD_FORK:
2185                 ret = machine__process_fork_event(machine, event, sample); break;
2186         case PERF_RECORD_EXIT:
2187                 ret = machine__process_exit_event(machine, event, sample); break;
2188         case PERF_RECORD_LOST:
2189                 ret = machine__process_lost_event(machine, event, sample); break;
2190         case PERF_RECORD_AUX:
2191                 ret = machine__process_aux_event(machine, event); break;
2192         case PERF_RECORD_ITRACE_START:
2193                 ret = machine__process_itrace_start_event(machine, event); break;
2194         case PERF_RECORD_LOST_SAMPLES:
2195                 ret = machine__process_lost_samples_event(machine, event, sample); break;
2196         case PERF_RECORD_SWITCH:
2197         case PERF_RECORD_SWITCH_CPU_WIDE:
2198                 ret = machine__process_switch_event(machine, event); break;
2199         case PERF_RECORD_KSYMBOL:
2200                 ret = machine__process_ksymbol(machine, event, sample); break;
2201         case PERF_RECORD_BPF_EVENT:
2202                 ret = machine__process_bpf(machine, event, sample); break;
2203         case PERF_RECORD_TEXT_POKE:
2204                 ret = machine__process_text_poke(machine, event, sample); break;
2205         case PERF_RECORD_AUX_OUTPUT_HW_ID:
2206                 ret = machine__process_aux_output_hw_id_event(machine, event); break;
2207         default:
2208                 ret = -1;
2209                 break;
2210         }
2211
2212         return ret;
2213 }
2214
2215 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
2216 {
2217         if (!regexec(regex, sym->name, 0, NULL, 0))
2218                 return true;
2219         return false;
2220 }
2221
2222 static void ip__resolve_ams(struct thread *thread,
2223                             struct addr_map_symbol *ams,
2224                             u64 ip)
2225 {
2226         struct addr_location al;
2227
2228         addr_location__init(&al);
2229         /*
2230          * We cannot use the header.misc hint to determine whether a
2231          * branch stack address is user, kernel, guest, hypervisor.
2232          * Branches may straddle the kernel/user/hypervisor boundaries.
2233          * Thus, we have to try consecutively until we find a match
2234          * or else, the symbol is unknown
2235          */
2236         thread__find_cpumode_addr_location(thread, ip, &al);
2237
2238         ams->addr = ip;
2239         ams->al_addr = al.addr;
2240         ams->al_level = al.level;
2241         ams->ms.maps = maps__get(al.maps);
2242         ams->ms.sym = al.sym;
2243         ams->ms.map = map__get(al.map);
2244         ams->phys_addr = 0;
2245         ams->data_page_size = 0;
2246         addr_location__exit(&al);
2247 }
2248
2249 static void ip__resolve_data(struct thread *thread,
2250                              u8 m, struct addr_map_symbol *ams,
2251                              u64 addr, u64 phys_addr, u64 daddr_page_size)
2252 {
2253         struct addr_location al;
2254
2255         addr_location__init(&al);
2256
2257         thread__find_symbol(thread, m, addr, &al);
2258
2259         ams->addr = addr;
2260         ams->al_addr = al.addr;
2261         ams->al_level = al.level;
2262         ams->ms.maps = maps__get(al.maps);
2263         ams->ms.sym = al.sym;
2264         ams->ms.map = map__get(al.map);
2265         ams->phys_addr = phys_addr;
2266         ams->data_page_size = daddr_page_size;
2267         addr_location__exit(&al);
2268 }
2269
2270 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
2271                                      struct addr_location *al)
2272 {
2273         struct mem_info *mi = mem_info__new();
2274
2275         if (!mi)
2276                 return NULL;
2277
2278         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
2279         ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2280                          sample->addr, sample->phys_addr,
2281                          sample->data_page_size);
2282         mi->data_src.val = sample->data_src;
2283
2284         return mi;
2285 }
2286
2287 static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2288 {
2289         struct map *map = ms->map;
2290         char *srcline = NULL;
2291         struct dso *dso;
2292
2293         if (!map || callchain_param.key == CCKEY_FUNCTION)
2294                 return srcline;
2295
2296         dso = map__dso(map);
2297         srcline = srcline__tree_find(&dso->srclines, ip);
2298         if (!srcline) {
2299                 bool show_sym = false;
2300                 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2301
2302                 srcline = get_srcline(dso, map__rip_2objdump(map, ip),
2303                                       ms->sym, show_sym, show_addr, ip);
2304                 srcline__tree_insert(&dso->srclines, ip, srcline);
2305         }
2306
2307         return srcline;
2308 }
2309
2310 struct iterations {
2311         int nr_loop_iter;
2312         u64 cycles;
2313 };
2314
2315 static int add_callchain_ip(struct thread *thread,
2316                             struct callchain_cursor *cursor,
2317                             struct symbol **parent,
2318                             struct addr_location *root_al,
2319                             u8 *cpumode,
2320                             u64 ip,
2321                             bool branch,
2322                             struct branch_flags *flags,
2323                             struct iterations *iter,
2324                             u64 branch_from)
2325 {
2326         struct map_symbol ms = {};
2327         struct addr_location al;
2328         int nr_loop_iter = 0, err = 0;
2329         u64 iter_cycles = 0;
2330         const char *srcline = NULL;
2331
2332         addr_location__init(&al);
2333         al.filtered = 0;
2334         al.sym = NULL;
2335         al.srcline = NULL;
2336         if (!cpumode) {
2337                 thread__find_cpumode_addr_location(thread, ip, &al);
2338         } else {
2339                 if (ip >= PERF_CONTEXT_MAX) {
2340                         switch (ip) {
2341                         case PERF_CONTEXT_HV:
2342                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2343                                 break;
2344                         case PERF_CONTEXT_KERNEL:
2345                                 *cpumode = PERF_RECORD_MISC_KERNEL;
2346                                 break;
2347                         case PERF_CONTEXT_USER:
2348                                 *cpumode = PERF_RECORD_MISC_USER;
2349                                 break;
2350                         default:
2351                                 pr_debug("invalid callchain context: "
2352                                          "%"PRId64"\n", (s64) ip);
2353                                 /*
2354                                  * It seems the callchain is corrupted.
2355                                  * Discard all.
2356                                  */
2357                                 callchain_cursor_reset(cursor);
2358                                 err = 1;
2359                                 goto out;
2360                         }
2361                         goto out;
2362                 }
2363                 thread__find_symbol(thread, *cpumode, ip, &al);
2364         }
2365
2366         if (al.sym != NULL) {
2367                 if (perf_hpp_list.parent && !*parent &&
2368                     symbol__match_regex(al.sym, &parent_regex))
2369                         *parent = al.sym;
2370                 else if (have_ignore_callees && root_al &&
2371                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
2372                         /* Treat this symbol as the root,
2373                            forgetting its callees. */
2374                         addr_location__copy(root_al, &al);
2375                         callchain_cursor_reset(cursor);
2376                 }
2377         }
2378
2379         if (symbol_conf.hide_unresolved && al.sym == NULL)
2380                 goto out;
2381
2382         if (iter) {
2383                 nr_loop_iter = iter->nr_loop_iter;
2384                 iter_cycles = iter->cycles;
2385         }
2386
2387         ms.maps = maps__get(al.maps);
2388         ms.map = map__get(al.map);
2389         ms.sym = al.sym;
2390
2391         if (!branch && append_inlines(cursor, &ms, ip) == 0)
2392                 goto out;
2393
2394         srcline = callchain_srcline(&ms, al.addr);
2395         err = callchain_cursor_append(cursor, ip, &ms,
2396                                       branch, flags, nr_loop_iter,
2397                                       iter_cycles, branch_from, srcline);
2398 out:
2399         addr_location__exit(&al);
2400         maps__put(ms.maps);
2401         map__put(ms.map);
2402         return err;
2403 }
2404
2405 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2406                                            struct addr_location *al)
2407 {
2408         unsigned int i;
2409         const struct branch_stack *bs = sample->branch_stack;
2410         struct branch_entry *entries = perf_sample__branch_entries(sample);
2411         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2412
2413         if (!bi)
2414                 return NULL;
2415
2416         for (i = 0; i < bs->nr; i++) {
2417                 ip__resolve_ams(al->thread, &bi[i].to, entries[i].to);
2418                 ip__resolve_ams(al->thread, &bi[i].from, entries[i].from);
2419                 bi[i].flags = entries[i].flags;
2420         }
2421         return bi;
2422 }
2423
2424 static void save_iterations(struct iterations *iter,
2425                             struct branch_entry *be, int nr)
2426 {
2427         int i;
2428
2429         iter->nr_loop_iter++;
2430         iter->cycles = 0;
2431
2432         for (i = 0; i < nr; i++)
2433                 iter->cycles += be[i].flags.cycles;
2434 }
2435
2436 #define CHASHSZ 127
2437 #define CHASHBITS 7
2438 #define NO_ENTRY 0xff
2439
2440 #define PERF_MAX_BRANCH_DEPTH 127
2441
2442 /* Remove loops. */
2443 static int remove_loops(struct branch_entry *l, int nr,
2444                         struct iterations *iter)
2445 {
2446         int i, j, off;
2447         unsigned char chash[CHASHSZ];
2448
2449         memset(chash, NO_ENTRY, sizeof(chash));
2450
2451         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2452
2453         for (i = 0; i < nr; i++) {
2454                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2455
2456                 /* no collision handling for now */
2457                 if (chash[h] == NO_ENTRY) {
2458                         chash[h] = i;
2459                 } else if (l[chash[h]].from == l[i].from) {
2460                         bool is_loop = true;
2461                         /* check if it is a real loop */
2462                         off = 0;
2463                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
2464                                 if (l[j].from != l[i + off].from) {
2465                                         is_loop = false;
2466                                         break;
2467                                 }
2468                         if (is_loop) {
2469                                 j = nr - (i + off);
2470                                 if (j > 0) {
2471                                         save_iterations(iter + i + off,
2472                                                 l + i, off);
2473
2474                                         memmove(iter + i, iter + i + off,
2475                                                 j * sizeof(*iter));
2476
2477                                         memmove(l + i, l + i + off,
2478                                                 j * sizeof(*l));
2479                                 }
2480
2481                                 nr -= off;
2482                         }
2483                 }
2484         }
2485         return nr;
2486 }
2487
2488 static int lbr_callchain_add_kernel_ip(struct thread *thread,
2489                                        struct callchain_cursor *cursor,
2490                                        struct perf_sample *sample,
2491                                        struct symbol **parent,
2492                                        struct addr_location *root_al,
2493                                        u64 branch_from,
2494                                        bool callee, int end)
2495 {
2496         struct ip_callchain *chain = sample->callchain;
2497         u8 cpumode = PERF_RECORD_MISC_USER;
2498         int err, i;
2499
2500         if (callee) {
2501                 for (i = 0; i < end + 1; i++) {
2502                         err = add_callchain_ip(thread, cursor, parent,
2503                                                root_al, &cpumode, chain->ips[i],
2504                                                false, NULL, NULL, branch_from);
2505                         if (err)
2506                                 return err;
2507                 }
2508                 return 0;
2509         }
2510
2511         for (i = end; i >= 0; i--) {
2512                 err = add_callchain_ip(thread, cursor, parent,
2513                                        root_al, &cpumode, chain->ips[i],
2514                                        false, NULL, NULL, branch_from);
2515                 if (err)
2516                         return err;
2517         }
2518
2519         return 0;
2520 }
2521
2522 static void save_lbr_cursor_node(struct thread *thread,
2523                                  struct callchain_cursor *cursor,
2524                                  int idx)
2525 {
2526         struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2527
2528         if (!lbr_stitch)
2529                 return;
2530
2531         if (cursor->pos == cursor->nr) {
2532                 lbr_stitch->prev_lbr_cursor[idx].valid = false;
2533                 return;
2534         }
2535
2536         if (!cursor->curr)
2537                 cursor->curr = cursor->first;
2538         else
2539                 cursor->curr = cursor->curr->next;
2540         memcpy(&lbr_stitch->prev_lbr_cursor[idx], cursor->curr,
2541                sizeof(struct callchain_cursor_node));
2542
2543         lbr_stitch->prev_lbr_cursor[idx].valid = true;
2544         cursor->pos++;
2545 }
2546
2547 static int lbr_callchain_add_lbr_ip(struct thread *thread,
2548                                     struct callchain_cursor *cursor,
2549                                     struct perf_sample *sample,
2550                                     struct symbol **parent,
2551                                     struct addr_location *root_al,
2552                                     u64 *branch_from,
2553                                     bool callee)
2554 {
2555         struct branch_stack *lbr_stack = sample->branch_stack;
2556         struct branch_entry *entries = perf_sample__branch_entries(sample);
2557         u8 cpumode = PERF_RECORD_MISC_USER;
2558         int lbr_nr = lbr_stack->nr;
2559         struct branch_flags *flags;
2560         int err, i;
2561         u64 ip;
2562
2563         /*
2564          * The curr and pos are not used in writing session. They are cleared
2565          * in callchain_cursor_commit() when the writing session is closed.
2566          * Using curr and pos to track the current cursor node.
2567          */
2568         if (thread__lbr_stitch(thread)) {
2569                 cursor->curr = NULL;
2570                 cursor->pos = cursor->nr;
2571                 if (cursor->nr) {
2572                         cursor->curr = cursor->first;
2573                         for (i = 0; i < (int)(cursor->nr - 1); i++)
2574                                 cursor->curr = cursor->curr->next;
2575                 }
2576         }
2577
2578         if (callee) {
2579                 /* Add LBR ip from first entries.to */
2580                 ip = entries[0].to;
2581                 flags = &entries[0].flags;
2582                 *branch_from = entries[0].from;
2583                 err = add_callchain_ip(thread, cursor, parent,
2584                                        root_al, &cpumode, ip,
2585                                        true, flags, NULL,
2586                                        *branch_from);
2587                 if (err)
2588                         return err;
2589
2590                 /*
2591                  * The number of cursor node increases.
2592                  * Move the current cursor node.
2593                  * But does not need to save current cursor node for entry 0.
2594                  * It's impossible to stitch the whole LBRs of previous sample.
2595                  */
2596                 if (thread__lbr_stitch(thread) && (cursor->pos != cursor->nr)) {
2597                         if (!cursor->curr)
2598                                 cursor->curr = cursor->first;
2599                         else
2600                                 cursor->curr = cursor->curr->next;
2601                         cursor->pos++;
2602                 }
2603
2604                 /* Add LBR ip from entries.from one by one. */
2605                 for (i = 0; i < lbr_nr; i++) {
2606                         ip = entries[i].from;
2607                         flags = &entries[i].flags;
2608                         err = add_callchain_ip(thread, cursor, parent,
2609                                                root_al, &cpumode, ip,
2610                                                true, flags, NULL,
2611                                                *branch_from);
2612                         if (err)
2613                                 return err;
2614                         save_lbr_cursor_node(thread, cursor, i);
2615                 }
2616                 return 0;
2617         }
2618
2619         /* Add LBR ip from entries.from one by one. */
2620         for (i = lbr_nr - 1; i >= 0; i--) {
2621                 ip = entries[i].from;
2622                 flags = &entries[i].flags;
2623                 err = add_callchain_ip(thread, cursor, parent,
2624                                        root_al, &cpumode, ip,
2625                                        true, flags, NULL,
2626                                        *branch_from);
2627                 if (err)
2628                         return err;
2629                 save_lbr_cursor_node(thread, cursor, i);
2630         }
2631
2632         /* Add LBR ip from first entries.to */
2633         ip = entries[0].to;
2634         flags = &entries[0].flags;
2635         *branch_from = entries[0].from;
2636         err = add_callchain_ip(thread, cursor, parent,
2637                                root_al, &cpumode, ip,
2638                                true, flags, NULL,
2639                                *branch_from);
2640         if (err)
2641                 return err;
2642
2643         return 0;
2644 }
2645
2646 static int lbr_callchain_add_stitched_lbr_ip(struct thread *thread,
2647                                              struct callchain_cursor *cursor)
2648 {
2649         struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2650         struct callchain_cursor_node *cnode;
2651         struct stitch_list *stitch_node;
2652         int err;
2653
2654         list_for_each_entry(stitch_node, &lbr_stitch->lists, node) {
2655                 cnode = &stitch_node->cursor;
2656
2657                 err = callchain_cursor_append(cursor, cnode->ip,
2658                                               &cnode->ms,
2659                                               cnode->branch,
2660                                               &cnode->branch_flags,
2661                                               cnode->nr_loop_iter,
2662                                               cnode->iter_cycles,
2663                                               cnode->branch_from,
2664                                               cnode->srcline);
2665                 if (err)
2666                         return err;
2667         }
2668         return 0;
2669 }
2670
2671 static struct stitch_list *get_stitch_node(struct thread *thread)
2672 {
2673         struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2674         struct stitch_list *stitch_node;
2675
2676         if (!list_empty(&lbr_stitch->free_lists)) {
2677                 stitch_node = list_first_entry(&lbr_stitch->free_lists,
2678                                                struct stitch_list, node);
2679                 list_del(&stitch_node->node);
2680
2681                 return stitch_node;
2682         }
2683
2684         return malloc(sizeof(struct stitch_list));
2685 }
2686
2687 static bool has_stitched_lbr(struct thread *thread,
2688                              struct perf_sample *cur,
2689                              struct perf_sample *prev,
2690                              unsigned int max_lbr,
2691                              bool callee)
2692 {
2693         struct branch_stack *cur_stack = cur->branch_stack;
2694         struct branch_entry *cur_entries = perf_sample__branch_entries(cur);
2695         struct branch_stack *prev_stack = prev->branch_stack;
2696         struct branch_entry *prev_entries = perf_sample__branch_entries(prev);
2697         struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2698         int i, j, nr_identical_branches = 0;
2699         struct stitch_list *stitch_node;
2700         u64 cur_base, distance;
2701
2702         if (!cur_stack || !prev_stack)
2703                 return false;
2704
2705         /* Find the physical index of the base-of-stack for current sample. */
2706         cur_base = max_lbr - cur_stack->nr + cur_stack->hw_idx + 1;
2707
2708         distance = (prev_stack->hw_idx > cur_base) ? (prev_stack->hw_idx - cur_base) :
2709                                                      (max_lbr + prev_stack->hw_idx - cur_base);
2710         /* Previous sample has shorter stack. Nothing can be stitched. */
2711         if (distance + 1 > prev_stack->nr)
2712                 return false;
2713
2714         /*
2715          * Check if there are identical LBRs between two samples.
2716          * Identical LBRs must have same from, to and flags values. Also,
2717          * they have to be saved in the same LBR registers (same physical
2718          * index).
2719          *
2720          * Starts from the base-of-stack of current sample.
2721          */
2722         for (i = distance, j = cur_stack->nr - 1; (i >= 0) && (j >= 0); i--, j--) {
2723                 if ((prev_entries[i].from != cur_entries[j].from) ||
2724                     (prev_entries[i].to != cur_entries[j].to) ||
2725                     (prev_entries[i].flags.value != cur_entries[j].flags.value))
2726                         break;
2727                 nr_identical_branches++;
2728         }
2729
2730         if (!nr_identical_branches)
2731                 return false;
2732
2733         /*
2734          * Save the LBRs between the base-of-stack of previous sample
2735          * and the base-of-stack of current sample into lbr_stitch->lists.
2736          * These LBRs will be stitched later.
2737          */
2738         for (i = prev_stack->nr - 1; i > (int)distance; i--) {
2739
2740                 if (!lbr_stitch->prev_lbr_cursor[i].valid)
2741                         continue;
2742
2743                 stitch_node = get_stitch_node(thread);
2744                 if (!stitch_node)
2745                         return false;
2746
2747                 memcpy(&stitch_node->cursor, &lbr_stitch->prev_lbr_cursor[i],
2748                        sizeof(struct callchain_cursor_node));
2749
2750                 if (callee)
2751                         list_add(&stitch_node->node, &lbr_stitch->lists);
2752                 else
2753                         list_add_tail(&stitch_node->node, &lbr_stitch->lists);
2754         }
2755
2756         return true;
2757 }
2758
2759 static bool alloc_lbr_stitch(struct thread *thread, unsigned int max_lbr)
2760 {
2761         if (thread__lbr_stitch(thread))
2762                 return true;
2763
2764         thread__set_lbr_stitch(thread, zalloc(sizeof(struct lbr_stitch)));
2765         if (!thread__lbr_stitch(thread))
2766                 goto err;
2767
2768         thread__lbr_stitch(thread)->prev_lbr_cursor =
2769                 calloc(max_lbr + 1, sizeof(struct callchain_cursor_node));
2770         if (!thread__lbr_stitch(thread)->prev_lbr_cursor)
2771                 goto free_lbr_stitch;
2772
2773         INIT_LIST_HEAD(&thread__lbr_stitch(thread)->lists);
2774         INIT_LIST_HEAD(&thread__lbr_stitch(thread)->free_lists);
2775
2776         return true;
2777
2778 free_lbr_stitch:
2779         free(thread__lbr_stitch(thread));
2780         thread__set_lbr_stitch(thread, NULL);
2781 err:
2782         pr_warning("Failed to allocate space for stitched LBRs. Disable LBR stitch\n");
2783         thread__set_lbr_stitch_enable(thread, false);
2784         return false;
2785 }
2786
2787 /*
2788  * Resolve LBR callstack chain sample
2789  * Return:
2790  * 1 on success get LBR callchain information
2791  * 0 no available LBR callchain information, should try fp
2792  * negative error code on other errors.
2793  */
2794 static int resolve_lbr_callchain_sample(struct thread *thread,
2795                                         struct callchain_cursor *cursor,
2796                                         struct perf_sample *sample,
2797                                         struct symbol **parent,
2798                                         struct addr_location *root_al,
2799                                         int max_stack,
2800                                         unsigned int max_lbr)
2801 {
2802         bool callee = (callchain_param.order == ORDER_CALLEE);
2803         struct ip_callchain *chain = sample->callchain;
2804         int chain_nr = min(max_stack, (int)chain->nr), i;
2805         struct lbr_stitch *lbr_stitch;
2806         bool stitched_lbr = false;
2807         u64 branch_from = 0;
2808         int err;
2809
2810         for (i = 0; i < chain_nr; i++) {
2811                 if (chain->ips[i] == PERF_CONTEXT_USER)
2812                         break;
2813         }
2814
2815         /* LBR only affects the user callchain */
2816         if (i == chain_nr)
2817                 return 0;
2818
2819         if (thread__lbr_stitch_enable(thread) && !sample->no_hw_idx &&
2820             (max_lbr > 0) && alloc_lbr_stitch(thread, max_lbr)) {
2821                 lbr_stitch = thread__lbr_stitch(thread);
2822
2823                 stitched_lbr = has_stitched_lbr(thread, sample,
2824                                                 &lbr_stitch->prev_sample,
2825                                                 max_lbr, callee);
2826
2827                 if (!stitched_lbr && !list_empty(&lbr_stitch->lists)) {
2828                         list_replace_init(&lbr_stitch->lists,
2829                                           &lbr_stitch->free_lists);
2830                 }
2831                 memcpy(&lbr_stitch->prev_sample, sample, sizeof(*sample));
2832         }
2833
2834         if (callee) {
2835                 /* Add kernel ip */
2836                 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2837                                                   parent, root_al, branch_from,
2838                                                   true, i);
2839                 if (err)
2840                         goto error;
2841
2842                 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2843                                                root_al, &branch_from, true);
2844                 if (err)
2845                         goto error;
2846
2847                 if (stitched_lbr) {
2848                         err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2849                         if (err)
2850                                 goto error;
2851                 }
2852
2853         } else {
2854                 if (stitched_lbr) {
2855                         err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2856                         if (err)
2857                                 goto error;
2858                 }
2859                 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2860                                                root_al, &branch_from, false);
2861                 if (err)
2862                         goto error;
2863
2864                 /* Add kernel ip */
2865                 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2866                                                   parent, root_al, branch_from,
2867                                                   false, i);
2868                 if (err)
2869                         goto error;
2870         }
2871         return 1;
2872
2873 error:
2874         return (err < 0) ? err : 0;
2875 }
2876
2877 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2878                              struct callchain_cursor *cursor,
2879                              struct symbol **parent,
2880                              struct addr_location *root_al,
2881                              u8 *cpumode, int ent)
2882 {
2883         int err = 0;
2884
2885         while (--ent >= 0) {
2886                 u64 ip = chain->ips[ent];
2887
2888                 if (ip >= PERF_CONTEXT_MAX) {
2889                         err = add_callchain_ip(thread, cursor, parent,
2890                                                root_al, cpumode, ip,
2891                                                false, NULL, NULL, 0);
2892                         break;
2893                 }
2894         }
2895         return err;
2896 }
2897
2898 static u64 get_leaf_frame_caller(struct perf_sample *sample,
2899                 struct thread *thread, int usr_idx)
2900 {
2901         if (machine__normalized_is(maps__machine(thread__maps(thread)), "arm64"))
2902                 return get_leaf_frame_caller_aarch64(sample, thread, usr_idx);
2903         else
2904                 return 0;
2905 }
2906
2907 static int thread__resolve_callchain_sample(struct thread *thread,
2908                                             struct callchain_cursor *cursor,
2909                                             struct evsel *evsel,
2910                                             struct perf_sample *sample,
2911                                             struct symbol **parent,
2912                                             struct addr_location *root_al,
2913                                             int max_stack)
2914 {
2915         struct branch_stack *branch = sample->branch_stack;
2916         struct branch_entry *entries = perf_sample__branch_entries(sample);
2917         struct ip_callchain *chain = sample->callchain;
2918         int chain_nr = 0;
2919         u8 cpumode = PERF_RECORD_MISC_USER;
2920         int i, j, err, nr_entries, usr_idx;
2921         int skip_idx = -1;
2922         int first_call = 0;
2923         u64 leaf_frame_caller;
2924
2925         if (chain)
2926                 chain_nr = chain->nr;
2927
2928         if (evsel__has_branch_callstack(evsel)) {
2929                 struct perf_env *env = evsel__env(evsel);
2930
2931                 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2932                                                    root_al, max_stack,
2933                                                    !env ? 0 : env->max_branches);
2934                 if (err)
2935                         return (err < 0) ? err : 0;
2936         }
2937
2938         /*
2939          * Based on DWARF debug information, some architectures skip
2940          * a callchain entry saved by the kernel.
2941          */
2942         skip_idx = arch_skip_callchain_idx(thread, chain);
2943
2944         /*
2945          * Add branches to call stack for easier browsing. This gives
2946          * more context for a sample than just the callers.
2947          *
2948          * This uses individual histograms of paths compared to the
2949          * aggregated histograms the normal LBR mode uses.
2950          *
2951          * Limitations for now:
2952          * - No extra filters
2953          * - No annotations (should annotate somehow)
2954          */
2955
2956         if (branch && callchain_param.branch_callstack) {
2957                 int nr = min(max_stack, (int)branch->nr);
2958                 struct branch_entry be[nr];
2959                 struct iterations iter[nr];
2960
2961                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2962                         pr_warning("corrupted branch chain. skipping...\n");
2963                         goto check_calls;
2964                 }
2965
2966                 for (i = 0; i < nr; i++) {
2967                         if (callchain_param.order == ORDER_CALLEE) {
2968                                 be[i] = entries[i];
2969
2970                                 if (chain == NULL)
2971                                         continue;
2972
2973                                 /*
2974                                  * Check for overlap into the callchain.
2975                                  * The return address is one off compared to
2976                                  * the branch entry. To adjust for this
2977                                  * assume the calling instruction is not longer
2978                                  * than 8 bytes.
2979                                  */
2980                                 if (i == skip_idx ||
2981                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
2982                                         first_call++;
2983                                 else if (be[i].from < chain->ips[first_call] &&
2984                                     be[i].from >= chain->ips[first_call] - 8)
2985                                         first_call++;
2986                         } else
2987                                 be[i] = entries[branch->nr - i - 1];
2988                 }
2989
2990                 memset(iter, 0, sizeof(struct iterations) * nr);
2991                 nr = remove_loops(be, nr, iter);
2992
2993                 for (i = 0; i < nr; i++) {
2994                         err = add_callchain_ip(thread, cursor, parent,
2995                                                root_al,
2996                                                NULL, be[i].to,
2997                                                true, &be[i].flags,
2998                                                NULL, be[i].from);
2999
3000                         if (!err)
3001                                 err = add_callchain_ip(thread, cursor, parent, root_al,
3002                                                        NULL, be[i].from,
3003                                                        true, &be[i].flags,
3004                                                        &iter[i], 0);
3005                         if (err == -EINVAL)
3006                                 break;
3007                         if (err)
3008                                 return err;
3009                 }
3010
3011                 if (chain_nr == 0)
3012                         return 0;
3013
3014                 chain_nr -= nr;
3015         }
3016
3017 check_calls:
3018         if (chain && callchain_param.order != ORDER_CALLEE) {
3019                 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
3020                                         &cpumode, chain->nr - first_call);
3021                 if (err)
3022                         return (err < 0) ? err : 0;
3023         }
3024         for (i = first_call, nr_entries = 0;
3025              i < chain_nr && nr_entries < max_stack; i++) {
3026                 u64 ip;
3027
3028                 if (callchain_param.order == ORDER_CALLEE)
3029                         j = i;
3030                 else
3031                         j = chain->nr - i - 1;
3032
3033 #ifdef HAVE_SKIP_CALLCHAIN_IDX
3034                 if (j == skip_idx)
3035                         continue;
3036 #endif
3037                 ip = chain->ips[j];
3038                 if (ip < PERF_CONTEXT_MAX)
3039                        ++nr_entries;
3040                 else if (callchain_param.order != ORDER_CALLEE) {
3041                         err = find_prev_cpumode(chain, thread, cursor, parent,
3042                                                 root_al, &cpumode, j);
3043                         if (err)
3044                                 return (err < 0) ? err : 0;
3045                         continue;
3046                 }
3047
3048                 /*
3049                  * PERF_CONTEXT_USER allows us to locate where the user stack ends.
3050                  * Depending on callchain_param.order and the position of PERF_CONTEXT_USER,
3051                  * the index will be different in order to add the missing frame
3052                  * at the right place.
3053                  */
3054
3055                 usr_idx = callchain_param.order == ORDER_CALLEE ? j-2 : j-1;
3056
3057                 if (usr_idx >= 0 && chain->ips[usr_idx] == PERF_CONTEXT_USER) {
3058
3059                         leaf_frame_caller = get_leaf_frame_caller(sample, thread, usr_idx);
3060
3061                         /*
3062                          * check if leaf_frame_Caller != ip to not add the same
3063                          * value twice.
3064                          */
3065
3066                         if (leaf_frame_caller && leaf_frame_caller != ip) {
3067
3068                                 err = add_callchain_ip(thread, cursor, parent,
3069                                                root_al, &cpumode, leaf_frame_caller,
3070                                                false, NULL, NULL, 0);
3071                                 if (err)
3072                                         return (err < 0) ? err : 0;
3073                         }
3074                 }
3075
3076                 err = add_callchain_ip(thread, cursor, parent,
3077                                        root_al, &cpumode, ip,
3078                                        false, NULL, NULL, 0);
3079
3080                 if (err)
3081                         return (err < 0) ? err : 0;
3082         }
3083
3084         return 0;
3085 }
3086
3087 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
3088 {
3089         struct symbol *sym = ms->sym;
3090         struct map *map = ms->map;
3091         struct inline_node *inline_node;
3092         struct inline_list *ilist;
3093         struct dso *dso;
3094         u64 addr;
3095         int ret = 1;
3096         struct map_symbol ilist_ms;
3097
3098         if (!symbol_conf.inline_name || !map || !sym)
3099                 return ret;
3100
3101         addr = map__dso_map_ip(map, ip);
3102         addr = map__rip_2objdump(map, addr);
3103         dso = map__dso(map);
3104
3105         inline_node = inlines__tree_find(&dso->inlined_nodes, addr);
3106         if (!inline_node) {
3107                 inline_node = dso__parse_addr_inlines(dso, addr, sym);
3108                 if (!inline_node)
3109                         return ret;
3110                 inlines__tree_insert(&dso->inlined_nodes, inline_node);
3111         }
3112
3113         ilist_ms = (struct map_symbol) {
3114                 .maps = maps__get(ms->maps),
3115                 .map = map__get(map),
3116         };
3117         list_for_each_entry(ilist, &inline_node->val, list) {
3118                 ilist_ms.sym = ilist->symbol;
3119                 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
3120                                               NULL, 0, 0, 0, ilist->srcline);
3121
3122                 if (ret != 0)
3123                         return ret;
3124         }
3125         map__put(ilist_ms.map);
3126         maps__put(ilist_ms.maps);
3127
3128         return ret;
3129 }
3130
3131 static int unwind_entry(struct unwind_entry *entry, void *arg)
3132 {
3133         struct callchain_cursor *cursor = arg;
3134         const char *srcline = NULL;
3135         u64 addr = entry->ip;
3136
3137         if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
3138                 return 0;
3139
3140         if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
3141                 return 0;
3142
3143         /*
3144          * Convert entry->ip from a virtual address to an offset in
3145          * its corresponding binary.
3146          */
3147         if (entry->ms.map)
3148                 addr = map__dso_map_ip(entry->ms.map, entry->ip);
3149
3150         srcline = callchain_srcline(&entry->ms, addr);
3151         return callchain_cursor_append(cursor, entry->ip, &entry->ms,
3152                                        false, NULL, 0, 0, 0, srcline);
3153 }
3154
3155 static int thread__resolve_callchain_unwind(struct thread *thread,
3156                                             struct callchain_cursor *cursor,
3157                                             struct evsel *evsel,
3158                                             struct perf_sample *sample,
3159                                             int max_stack)
3160 {
3161         /* Can we do dwarf post unwind? */
3162         if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
3163               (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
3164                 return 0;
3165
3166         /* Bail out if nothing was captured. */
3167         if ((!sample->user_regs.regs) ||
3168             (!sample->user_stack.size))
3169                 return 0;
3170
3171         return unwind__get_entries(unwind_entry, cursor,
3172                                    thread, sample, max_stack, false);
3173 }
3174
3175 int thread__resolve_callchain(struct thread *thread,
3176                               struct callchain_cursor *cursor,
3177                               struct evsel *evsel,
3178                               struct perf_sample *sample,
3179                               struct symbol **parent,
3180                               struct addr_location *root_al,
3181                               int max_stack)
3182 {
3183         int ret = 0;
3184
3185         if (cursor == NULL)
3186                 return -ENOMEM;
3187
3188         callchain_cursor_reset(cursor);
3189
3190         if (callchain_param.order == ORDER_CALLEE) {
3191                 ret = thread__resolve_callchain_sample(thread, cursor,
3192                                                        evsel, sample,
3193                                                        parent, root_al,
3194                                                        max_stack);
3195                 if (ret)
3196                         return ret;
3197                 ret = thread__resolve_callchain_unwind(thread, cursor,
3198                                                        evsel, sample,
3199                                                        max_stack);
3200         } else {
3201                 ret = thread__resolve_callchain_unwind(thread, cursor,
3202                                                        evsel, sample,
3203                                                        max_stack);
3204                 if (ret)
3205                         return ret;
3206                 ret = thread__resolve_callchain_sample(thread, cursor,
3207                                                        evsel, sample,
3208                                                        parent, root_al,
3209                                                        max_stack);
3210         }
3211
3212         return ret;
3213 }
3214
3215 int machine__for_each_thread(struct machine *machine,
3216                              int (*fn)(struct thread *thread, void *p),
3217                              void *priv)
3218 {
3219         struct threads *threads;
3220         struct rb_node *nd;
3221         int rc = 0;
3222         int i;
3223
3224         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3225                 threads = &machine->threads[i];
3226                 for (nd = rb_first_cached(&threads->entries); nd;
3227                      nd = rb_next(nd)) {
3228                         struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
3229
3230                         rc = fn(trb->thread, priv);
3231                         if (rc != 0)
3232                                 return rc;
3233                 }
3234         }
3235         return rc;
3236 }
3237
3238 int machines__for_each_thread(struct machines *machines,
3239                               int (*fn)(struct thread *thread, void *p),
3240                               void *priv)
3241 {
3242         struct rb_node *nd;
3243         int rc = 0;
3244
3245         rc = machine__for_each_thread(&machines->host, fn, priv);
3246         if (rc != 0)
3247                 return rc;
3248
3249         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
3250                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
3251
3252                 rc = machine__for_each_thread(machine, fn, priv);
3253                 if (rc != 0)
3254                         return rc;
3255         }
3256         return rc;
3257 }
3258
3259 pid_t machine__get_current_tid(struct machine *machine, int cpu)
3260 {
3261         if (cpu < 0 || (size_t)cpu >= machine->current_tid_sz)
3262                 return -1;
3263
3264         return machine->current_tid[cpu];
3265 }
3266
3267 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
3268                              pid_t tid)
3269 {
3270         struct thread *thread;
3271         const pid_t init_val = -1;
3272
3273         if (cpu < 0)
3274                 return -EINVAL;
3275
3276         if (realloc_array_as_needed(machine->current_tid,
3277                                     machine->current_tid_sz,
3278                                     (unsigned int)cpu,
3279                                     &init_val))
3280                 return -ENOMEM;
3281
3282         machine->current_tid[cpu] = tid;
3283
3284         thread = machine__findnew_thread(machine, pid, tid);
3285         if (!thread)
3286                 return -ENOMEM;
3287
3288         thread__set_cpu(thread, cpu);
3289         thread__put(thread);
3290
3291         return 0;
3292 }
3293
3294 /*
3295  * Compares the raw arch string. N.B. see instead perf_env__arch() or
3296  * machine__normalized_is() if a normalized arch is needed.
3297  */
3298 bool machine__is(struct machine *machine, const char *arch)
3299 {
3300         return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
3301 }
3302
3303 bool machine__normalized_is(struct machine *machine, const char *arch)
3304 {
3305         return machine && !strcmp(perf_env__arch(machine->env), arch);
3306 }
3307
3308 int machine__nr_cpus_avail(struct machine *machine)
3309 {
3310         return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
3311 }
3312
3313 int machine__get_kernel_start(struct machine *machine)
3314 {
3315         struct map *map = machine__kernel_map(machine);
3316         int err = 0;
3317
3318         /*
3319          * The only addresses above 2^63 are kernel addresses of a 64-bit
3320          * kernel.  Note that addresses are unsigned so that on a 32-bit system
3321          * all addresses including kernel addresses are less than 2^32.  In
3322          * that case (32-bit system), if the kernel mapping is unknown, all
3323          * addresses will be assumed to be in user space - see
3324          * machine__kernel_ip().
3325          */
3326         machine->kernel_start = 1ULL << 63;
3327         if (map) {
3328                 err = map__load(map);
3329                 /*
3330                  * On x86_64, PTI entry trampolines are less than the
3331                  * start of kernel text, but still above 2^63. So leave
3332                  * kernel_start = 1ULL << 63 for x86_64.
3333                  */
3334                 if (!err && !machine__is(machine, "x86_64"))
3335                         machine->kernel_start = map__start(map);
3336         }
3337         return err;
3338 }
3339
3340 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
3341 {
3342         u8 addr_cpumode = cpumode;
3343         bool kernel_ip;
3344
3345         if (!machine->single_address_space)
3346                 goto out;
3347
3348         kernel_ip = machine__kernel_ip(machine, addr);
3349         switch (cpumode) {
3350         case PERF_RECORD_MISC_KERNEL:
3351         case PERF_RECORD_MISC_USER:
3352                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
3353                                            PERF_RECORD_MISC_USER;
3354                 break;
3355         case PERF_RECORD_MISC_GUEST_KERNEL:
3356         case PERF_RECORD_MISC_GUEST_USER:
3357                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
3358                                            PERF_RECORD_MISC_GUEST_USER;
3359                 break;
3360         default:
3361                 break;
3362         }
3363 out:
3364         return addr_cpumode;
3365 }
3366
3367 struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id)
3368 {
3369         return dsos__findnew_id(&machine->dsos, filename, id);
3370 }
3371
3372 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
3373 {
3374         return machine__findnew_dso_id(machine, filename, NULL);
3375 }
3376
3377 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
3378 {
3379         struct machine *machine = vmachine;
3380         struct map *map;
3381         struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
3382
3383         if (sym == NULL)
3384                 return NULL;
3385
3386         *modp = __map__is_kmodule(map) ? (char *)map__dso(map)->short_name : NULL;
3387         *addrp = map__unmap_ip(map, sym->start);
3388         return sym->name;
3389 }
3390
3391 int machine__for_each_dso(struct machine *machine, machine__dso_t fn, void *priv)
3392 {
3393         struct dso *pos;
3394         int err = 0;
3395
3396         list_for_each_entry(pos, &machine->dsos.head, node) {
3397                 if (fn(pos, machine, priv))
3398                         err = -1;
3399         }
3400         return err;
3401 }
3402
3403 int machine__for_each_kernel_map(struct machine *machine, machine__map_t fn, void *priv)
3404 {
3405         struct maps *maps = machine__kernel_maps(machine);
3406         struct map_rb_node *pos;
3407         int err = 0;
3408
3409         maps__for_each_entry(maps, pos) {
3410                 err = fn(pos->map, priv);
3411                 if (err != 0) {
3412                         break;
3413                 }
3414         }
3415         return err;
3416 }
3417
3418 bool machine__is_lock_function(struct machine *machine, u64 addr)
3419 {
3420         if (!machine->sched.text_start) {
3421                 struct map *kmap;
3422                 struct symbol *sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_start", &kmap);
3423
3424                 if (!sym) {
3425                         /* to avoid retry */
3426                         machine->sched.text_start = 1;
3427                         return false;
3428                 }
3429
3430                 machine->sched.text_start = map__unmap_ip(kmap, sym->start);
3431
3432                 /* should not fail from here */
3433                 sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_end", &kmap);
3434                 machine->sched.text_end = map__unmap_ip(kmap, sym->start);
3435
3436                 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_start", &kmap);
3437                 machine->lock.text_start = map__unmap_ip(kmap, sym->start);
3438
3439                 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_end", &kmap);
3440                 machine->lock.text_end = map__unmap_ip(kmap, sym->start);
3441         }
3442
3443         /* failed to get kernel symbols */
3444         if (machine->sched.text_start == 1)
3445                 return false;
3446
3447         /* mutex and rwsem functions are in sched text section */
3448         if (machine->sched.text_start <= addr && addr < machine->sched.text_end)
3449                 return true;
3450
3451         /* spinlock functions are in lock text section */
3452         if (machine->lock.text_start <= addr && addr < machine->lock.text_end)
3453                 return true;
3454
3455         return false;
3456 }