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