fs/btrfs/ref-verify.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2014 Facebook.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/stacktrace.h>
   8 #include "messages.h"
   9 #include "ctree.h"
  10 #include "disk-io.h"
  11 #include "locking.h"
  12 #include "delayed-ref.h"
  13 #include "ref-verify.h"
  14 #include "fs.h"
  15 #include "accessors.h"
  16
  17 /*
  18  * Used to keep track the roots and number of refs each root has for a given
  19  * bytenr.  This just tracks the number of direct references, no shared
  20  * references.
  21  */
  22 struct root_entry {
  23         u64 root_objectid;
  24         u64 num_refs;
  25         struct rb_node node;
  26 };
  27
  28 /*
  29  * These are meant to represent what should exist in the extent tree, these can
  30  * be used to verify the extent tree is consistent as these should all match
  31  * what the extent tree says.
  32  */
  33 struct ref_entry {
  34         u64 root_objectid;
  35         u64 parent;
  36         u64 owner;
  37         u64 offset;
  38         u64 num_refs;
  39         struct rb_node node;
  40 };
  41
  42 #define MAX_TRACE       16
  43
  44 /*
  45  * Whenever we add/remove a reference we record the action.  The action maps
  46  * back to the delayed ref action.  We hold the ref we are changing in the
  47  * action so we can account for the history properly, and we record the root we
  48  * were called with since it could be different from ref_root.  We also store
  49  * stack traces because that's how I roll.
  50  */
  51 struct ref_action {
  52         int action;
  53         u64 root;
  54         struct ref_entry ref;
  55         struct list_head list;
  56         unsigned long trace[MAX_TRACE];
  57         unsigned int trace_len;
  58 };
  59
  60 /*
  61  * One of these for every block we reference, it holds the roots and references
  62  * to it as well as all of the ref actions that have occurred to it.  We never
  63  * free it until we unmount the file system in order to make sure re-allocations
  64  * are happening properly.
  65  */
  66 struct block_entry {
  67         u64 bytenr;
  68         u64 len;
  69         u64 num_refs;
  70         int metadata;
  71         int from_disk;
  72         struct rb_root roots;
  73         struct rb_root refs;
  74         struct rb_node node;
  75         struct list_head actions;
  76 };
  77
  78 static struct block_entry *insert_block_entry(struct rb_root *root,
  79                                               struct block_entry *be)
  80 {
  81         struct rb_node **p = &root->rb_node;
  82         struct rb_node *parent_node = NULL;
  83         struct block_entry *entry;
  84
  85         while (*p) {
  86                 parent_node = *p;
  87                 entry = rb_entry(parent_node, struct block_entry, node);
  88                 if (entry->bytenr > be->bytenr)
  89                         p = &(*p)->rb_left;
  90                 else if (entry->bytenr < be->bytenr)
  91                         p = &(*p)->rb_right;
  92                 else
  93                         return entry;
  94         }
  95
  96         rb_link_node(&be->node, parent_node, p);
  97         rb_insert_color(&be->node, root);
  98         return NULL;
  99 }
 100
 101 static struct block_entry *lookup_block_entry(struct rb_root *root, u64 bytenr)
 102 {
 103         struct rb_node *n;
 104         struct block_entry *entry = NULL;
 105
 106         n = root->rb_node;
 107         while (n) {
 108                 entry = rb_entry(n, struct block_entry, node);
 109                 if (entry->bytenr < bytenr)
 110                         n = n->rb_right;
 111                 else if (entry->bytenr > bytenr)
 112                         n = n->rb_left;
 113                 else
 114                         return entry;
 115         }
 116         return NULL;
 117 }
 118
 119 static struct root_entry *insert_root_entry(struct rb_root *root,
 120                                             struct root_entry *re)
 121 {
 122         struct rb_node **p = &root->rb_node;
 123         struct rb_node *parent_node = NULL;
 124         struct root_entry *entry;
 125
 126         while (*p) {
 127                 parent_node = *p;
 128                 entry = rb_entry(parent_node, struct root_entry, node);
 129                 if (entry->root_objectid > re->root_objectid)
 130                         p = &(*p)->rb_left;
 131                 else if (entry->root_objectid < re->root_objectid)
 132                         p = &(*p)->rb_right;
 133                 else
 134                         return entry;
 135         }
 136
 137         rb_link_node(&re->node, parent_node, p);
 138         rb_insert_color(&re->node, root);
 139         return NULL;
 140
 141 }
 142
 143 static int comp_refs(struct ref_entry *ref1, struct ref_entry *ref2)
 144 {
 145         if (ref1->root_objectid < ref2->root_objectid)
 146                 return -1;
 147         if (ref1->root_objectid > ref2->root_objectid)
 148                 return 1;
 149         if (ref1->parent < ref2->parent)
 150                 return -1;
 151         if (ref1->parent > ref2->parent)
 152                 return 1;
 153         if (ref1->owner < ref2->owner)
 154                 return -1;
 155         if (ref1->owner > ref2->owner)
 156                 return 1;
 157         if (ref1->offset < ref2->offset)
 158                 return -1;
 159         if (ref1->offset > ref2->offset)
 160                 return 1;
 161         return 0;
 162 }
 163
 164 static struct ref_entry *insert_ref_entry(struct rb_root *root,
 165                                           struct ref_entry *ref)
 166 {
 167         struct rb_node **p = &root->rb_node;
 168         struct rb_node *parent_node = NULL;
 169         struct ref_entry *entry;
 170         int cmp;
 171
 172         while (*p) {
 173                 parent_node = *p;
 174                 entry = rb_entry(parent_node, struct ref_entry, node);
 175                 cmp = comp_refs(entry, ref);
 176                 if (cmp > 0)
 177                         p = &(*p)->rb_left;
 178                 else if (cmp < 0)
 179                         p = &(*p)->rb_right;
 180                 else
 181                         return entry;
 182         }
 183
 184         rb_link_node(&ref->node, parent_node, p);
 185         rb_insert_color(&ref->node, root);
 186         return NULL;
 187
 188 }
 189
 190 static struct root_entry *lookup_root_entry(struct rb_root *root, u64 objectid)
 191 {
 192         struct rb_node *n;
 193         struct root_entry *entry = NULL;
 194
 195         n = root->rb_node;
 196         while (n) {
 197                 entry = rb_entry(n, struct root_entry, node);
 198                 if (entry->root_objectid < objectid)
 199                         n = n->rb_right;
 200                 else if (entry->root_objectid > objectid)
 201                         n = n->rb_left;
 202                 else
 203                         return entry;
 204         }
 205         return NULL;
 206 }
 207
 208 #ifdef CONFIG_STACKTRACE
 209 static void __save_stack_trace(struct ref_action *ra)
 210 {
 211         ra->trace_len = stack_trace_save(ra->trace, MAX_TRACE, 2);
 212 }
 213
 214 static void __print_stack_trace(struct btrfs_fs_info *fs_info,
 215                                 struct ref_action *ra)
 216 {
 217         if (ra->trace_len == 0) {
 218                 btrfs_err(fs_info, "  ref-verify: no stacktrace");
 219                 return;
 220         }
 221         stack_trace_print(ra->trace, ra->trace_len, 2);
 222 }
 223 #else
 224 static inline void __save_stack_trace(struct ref_action *ra)
 225 {
 226 }
 227
 228 static inline void __print_stack_trace(struct btrfs_fs_info *fs_info,
 229                                        struct ref_action *ra)
 230 {
 231         btrfs_err(fs_info, "  ref-verify: no stacktrace support");
 232 }
 233 #endif
 234
 235 static void free_block_entry(struct block_entry *be)
 236 {
 237         struct root_entry *re;
 238         struct ref_entry *ref;
 239         struct ref_action *ra;
 240         struct rb_node *n;
 241
 242         while ((n = rb_first(&be->roots))) {
 243                 re = rb_entry(n, struct root_entry, node);
 244                 rb_erase(&re->node, &be->roots);
 245                 kfree(re);
 246         }
 247
 248         while((n = rb_first(&be->refs))) {
 249                 ref = rb_entry(n, struct ref_entry, node);
 250                 rb_erase(&ref->node, &be->refs);
 251                 kfree(ref);
 252         }
 253
 254         while (!list_empty(&be->actions)) {
 255                 ra = list_first_entry(&be->actions, struct ref_action,
 256                                       list);
 257                 list_del(&ra->list);
 258                 kfree(ra);
 259         }
 260         kfree(be);
 261 }
 262
 263 static struct block_entry *add_block_entry(struct btrfs_fs_info *fs_info,
 264                                            u64 bytenr, u64 len,
 265                                            u64 root_objectid)
 266 {
 267         struct block_entry *be = NULL, *exist;
 268         struct root_entry *re = NULL;
 269
 270         re = kzalloc(sizeof(struct root_entry), GFP_NOFS);
 271         be = kzalloc(sizeof(struct block_entry), GFP_NOFS);
 272         if (!be || !re) {
 273                 kfree(re);
 274                 kfree(be);
 275                 return ERR_PTR(-ENOMEM);
 276         }
 277         be->bytenr = bytenr;
 278         be->len = len;
 279
 280         re->root_objectid = root_objectid;
 281         re->num_refs = 0;
 282
 283         spin_lock(&fs_info->ref_verify_lock);
 284         exist = insert_block_entry(&fs_info->block_tree, be);
 285         if (exist) {
 286                 if (root_objectid) {
 287                         struct root_entry *exist_re;
 288
 289                         exist_re = insert_root_entry(&exist->roots, re);
 290                         if (exist_re)
 291                                 kfree(re);
 292                 } else {
 293                         kfree(re);
 294                 }
 295                 kfree(be);
 296                 return exist;
 297         }
 298
 299         be->num_refs = 0;
 300         be->metadata = 0;
 301         be->from_disk = 0;
 302         be->roots = RB_ROOT;
 303         be->refs = RB_ROOT;
 304         INIT_LIST_HEAD(&be->actions);
 305         if (root_objectid)
 306                 insert_root_entry(&be->roots, re);
 307         else
 308                 kfree(re);
 309         return be;
 310 }
 311
 312 static int add_tree_block(struct btrfs_fs_info *fs_info, u64 ref_root,
 313                           u64 parent, u64 bytenr, int level)
 314 {
 315         struct block_entry *be;
 316         struct root_entry *re;
 317         struct ref_entry *ref = NULL, *exist;
 318
 319         ref = kmalloc(sizeof(struct ref_entry), GFP_NOFS);
 320         if (!ref)
 321                 return -ENOMEM;
 322
 323         if (parent)
 324                 ref->root_objectid = 0;
 325         else
 326                 ref->root_objectid = ref_root;
 327         ref->parent = parent;
 328         ref->owner = level;
 329         ref->offset = 0;
 330         ref->num_refs = 1;
 331
 332         be = add_block_entry(fs_info, bytenr, fs_info->nodesize, ref_root);
 333         if (IS_ERR(be)) {
 334                 kfree(ref);
 335                 return PTR_ERR(be);
 336         }
 337         be->num_refs++;
 338         be->from_disk = 1;
 339         be->metadata = 1;
 340
 341         if (!parent) {
 342                 ASSERT(ref_root);
 343                 re = lookup_root_entry(&be->roots, ref_root);
 344                 ASSERT(re);
 345                 re->num_refs++;
 346         }
 347         exist = insert_ref_entry(&be->refs, ref);
 348         if (exist) {
 349                 exist->num_refs++;
 350                 kfree(ref);
 351         }
 352         spin_unlock(&fs_info->ref_verify_lock);
 353
 354         return 0;
 355 }
 356
 357 static int add_shared_data_ref(struct btrfs_fs_info *fs_info,
 358                                u64 parent, u32 num_refs, u64 bytenr,
 359                                u64 num_bytes)
 360 {
 361         struct block_entry *be;
 362         struct ref_entry *ref;
 363
 364         ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
 365         if (!ref)
 366                 return -ENOMEM;
 367         be = add_block_entry(fs_info, bytenr, num_bytes, 0);
 368         if (IS_ERR(be)) {
 369                 kfree(ref);
 370                 return PTR_ERR(be);
 371         }
 372         be->num_refs += num_refs;
 373
 374         ref->parent = parent;
 375         ref->num_refs = num_refs;
 376         if (insert_ref_entry(&be->refs, ref)) {
 377                 spin_unlock(&fs_info->ref_verify_lock);
 378                 btrfs_err(fs_info, "existing shared ref when reading from disk?");
 379                 kfree(ref);
 380                 return -EINVAL;
 381         }
 382         spin_unlock(&fs_info->ref_verify_lock);
 383         return 0;
 384 }
 385
 386 static int add_extent_data_ref(struct btrfs_fs_info *fs_info,
 387                                struct extent_buffer *leaf,
 388                                struct btrfs_extent_data_ref *dref,
 389                                u64 bytenr, u64 num_bytes)
 390 {
 391         struct block_entry *be;
 392         struct ref_entry *ref;
 393         struct root_entry *re;
 394         u64 ref_root = btrfs_extent_data_ref_root(leaf, dref);
 395         u64 owner = btrfs_extent_data_ref_objectid(leaf, dref);
 396         u64 offset = btrfs_extent_data_ref_offset(leaf, dref);
 397         u32 num_refs = btrfs_extent_data_ref_count(leaf, dref);
 398
 399         ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
 400         if (!ref)
 401                 return -ENOMEM;
 402         be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
 403         if (IS_ERR(be)) {
 404                 kfree(ref);
 405                 return PTR_ERR(be);
 406         }
 407         be->num_refs += num_refs;
 408
 409         ref->parent = 0;
 410         ref->owner = owner;
 411         ref->root_objectid = ref_root;
 412         ref->offset = offset;
 413         ref->num_refs = num_refs;
 414         if (insert_ref_entry(&be->refs, ref)) {
 415                 spin_unlock(&fs_info->ref_verify_lock);
 416                 btrfs_err(fs_info, "existing ref when reading from disk?");
 417                 kfree(ref);
 418                 return -EINVAL;
 419         }
 420
 421         re = lookup_root_entry(&be->roots, ref_root);
 422         if (!re) {
 423                 spin_unlock(&fs_info->ref_verify_lock);
 424                 btrfs_err(fs_info, "missing root in new block entry?");
 425                 return -EINVAL;
 426         }
 427         re->num_refs += num_refs;
 428         spin_unlock(&fs_info->ref_verify_lock);
 429         return 0;
 430 }
 431
 432 static int process_extent_item(struct btrfs_fs_info *fs_info,
 433                                struct btrfs_path *path, struct btrfs_key *key,
 434                                int slot, int *tree_block_level)
 435 {
 436         struct btrfs_extent_item *ei;
 437         struct btrfs_extent_inline_ref *iref;
 438         struct btrfs_extent_data_ref *dref;
 439         struct btrfs_shared_data_ref *sref;
 440         struct extent_buffer *leaf = path->nodes[0];
 441         u32 item_size = btrfs_item_size(leaf, slot);
 442         unsigned long end, ptr;
 443         u64 offset, flags, count;
 444         int type, ret;
 445
 446         ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item);
 447         flags = btrfs_extent_flags(leaf, ei);
 448
 449         if ((key->type == BTRFS_EXTENT_ITEM_KEY) &&
 450             flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 451                 struct btrfs_tree_block_info *info;
 452
 453                 info = (struct btrfs_tree_block_info *)(ei + 1);
 454                 *tree_block_level = btrfs_tree_block_level(leaf, info);
 455                 iref = (struct btrfs_extent_inline_ref *)(info + 1);
 456         } else {
 457                 if (key->type == BTRFS_METADATA_ITEM_KEY)
 458                         *tree_block_level = key->offset;
 459                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
 460         }
 461
 462         ptr = (unsigned long)iref;
 463         end = (unsigned long)ei + item_size;
 464         while (ptr < end) {
 465                 iref = (struct btrfs_extent_inline_ref *)ptr;
 466                 type = btrfs_extent_inline_ref_type(leaf, iref);
 467                 offset = btrfs_extent_inline_ref_offset(leaf, iref);
 468                 switch (type) {
 469                 case BTRFS_TREE_BLOCK_REF_KEY:
 470                         ret = add_tree_block(fs_info, offset, 0, key->objectid,
 471                                              *tree_block_level);
 472                         break;
 473                 case BTRFS_SHARED_BLOCK_REF_KEY:
 474                         ret = add_tree_block(fs_info, 0, offset, key->objectid,
 475                                              *tree_block_level);
 476                         break;
 477                 case BTRFS_EXTENT_DATA_REF_KEY:
 478                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
 479                         ret = add_extent_data_ref(fs_info, leaf, dref,
 480                                                   key->objectid, key->offset);
 481                         break;
 482                 case BTRFS_SHARED_DATA_REF_KEY:
 483                         sref = (struct btrfs_shared_data_ref *)(iref + 1);
 484                         count = btrfs_shared_data_ref_count(leaf, sref);
 485                         ret = add_shared_data_ref(fs_info, offset, count,
 486                                                   key->objectid, key->offset);
 487                         break;
 488                 case BTRFS_EXTENT_OWNER_REF_KEY:
 489                         WARN_ON(!btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 490                         break;
 491                 default:
 492                         btrfs_err(fs_info, "invalid key type in iref");
 493                         ret = -EINVAL;
 494                         break;
 495                 }
 496                 if (ret)
 497                         break;
 498                 ptr += btrfs_extent_inline_ref_size(type);
 499         }
 500         return ret;
 501 }
 502
 503 static int process_leaf(struct btrfs_root *root,
 504                         struct btrfs_path *path, u64 *bytenr, u64 *num_bytes,
 505                         int *tree_block_level)
 506 {
 507         struct btrfs_fs_info *fs_info = root->fs_info;
 508         struct extent_buffer *leaf = path->nodes[0];
 509         struct btrfs_extent_data_ref *dref;
 510         struct btrfs_shared_data_ref *sref;
 511         u32 count;
 512         int i = 0, ret = 0;
 513         struct btrfs_key key;
 514         int nritems = btrfs_header_nritems(leaf);
 515
 516         for (i = 0; i < nritems; i++) {
 517                 btrfs_item_key_to_cpu(leaf, &key, i);
 518                 switch (key.type) {
 519                 case BTRFS_EXTENT_ITEM_KEY:
 520                         *num_bytes = key.offset;
 521                         fallthrough;
 522                 case BTRFS_METADATA_ITEM_KEY:
 523                         *bytenr = key.objectid;
 524                         ret = process_extent_item(fs_info, path, &key, i,
 525                                                   tree_block_level);
 526                         break;
 527                 case BTRFS_TREE_BLOCK_REF_KEY:
 528                         ret = add_tree_block(fs_info, key.offset, 0,
 529                                              key.objectid, *tree_block_level);
 530                         break;
 531                 case BTRFS_SHARED_BLOCK_REF_KEY:
 532                         ret = add_tree_block(fs_info, 0, key.offset,
 533                                              key.objectid, *tree_block_level);
 534                         break;
 535                 case BTRFS_EXTENT_DATA_REF_KEY:
 536                         dref = btrfs_item_ptr(leaf, i,
 537                                               struct btrfs_extent_data_ref);
 538                         ret = add_extent_data_ref(fs_info, leaf, dref, *bytenr,
 539                                                   *num_bytes);
 540                         break;
 541                 case BTRFS_SHARED_DATA_REF_KEY:
 542                         sref = btrfs_item_ptr(leaf, i,
 543                                               struct btrfs_shared_data_ref);
 544                         count = btrfs_shared_data_ref_count(leaf, sref);
 545                         ret = add_shared_data_ref(fs_info, key.offset, count,
 546                                                   *bytenr, *num_bytes);
 547                         break;
 548                 default:
 549                         break;
 550                 }
 551                 if (ret)
 552                         break;
 553         }
 554         return ret;
 555 }
 556
 557 /* Walk down to the leaf from the given level */
 558 static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
 559                           int level, u64 *bytenr, u64 *num_bytes,
 560                           int *tree_block_level)
 561 {
 562         struct extent_buffer *eb;
 563         int ret = 0;
 564
 565         while (level >= 0) {
 566                 if (level) {
 567                         eb = btrfs_read_node_slot(path->nodes[level],
 568                                                   path->slots[level]);
 569                         if (IS_ERR(eb))
 570                                 return PTR_ERR(eb);
 571                         btrfs_tree_read_lock(eb);
 572                         path->nodes[level-1] = eb;
 573                         path->slots[level-1] = 0;
 574                         path->locks[level-1] = BTRFS_READ_LOCK;
 575                 } else {
 576                         ret = process_leaf(root, path, bytenr, num_bytes,
 577                                            tree_block_level);
 578                         if (ret)
 579                                 break;
 580                 }
 581                 level--;
 582         }
 583         return ret;
 584 }
 585
 586 /* Walk up to the next node that needs to be processed */
 587 static int walk_up_tree(struct btrfs_path *path, int *level)
 588 {
 589         int l;
 590
 591         for (l = 0; l < BTRFS_MAX_LEVEL; l++) {
 592                 if (!path->nodes[l])
 593                         continue;
 594                 if (l) {
 595                         path->slots[l]++;
 596                         if (path->slots[l] <
 597                             btrfs_header_nritems(path->nodes[l])) {
 598                                 *level = l;
 599                                 return 0;
 600                         }
 601                 }
 602                 btrfs_tree_unlock_rw(path->nodes[l], path->locks[l]);
 603                 free_extent_buffer(path->nodes[l]);
 604                 path->nodes[l] = NULL;
 605                 path->slots[l] = 0;
 606                 path->locks[l] = 0;
 607         }
 608
 609         return 1;
 610 }
 611
 612 static void dump_ref_action(struct btrfs_fs_info *fs_info,
 613                             struct ref_action *ra)
 614 {
 615         btrfs_err(fs_info,
 616 "  Ref action %d, root %llu, ref_root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
 617                   ra->action, ra->root, ra->ref.root_objectid, ra->ref.parent,
 618                   ra->ref.owner, ra->ref.offset, ra->ref.num_refs);
 619         __print_stack_trace(fs_info, ra);
 620 }
 621
 622 /*
 623  * Dumps all the information from the block entry to printk, it's going to be
 624  * awesome.
 625  */
 626 static void dump_block_entry(struct btrfs_fs_info *fs_info,
 627                              struct block_entry *be)
 628 {
 629         struct ref_entry *ref;
 630         struct root_entry *re;
 631         struct ref_action *ra;
 632         struct rb_node *n;
 633
 634         btrfs_err(fs_info,
 635 "dumping block entry [%llu %llu], num_refs %llu, metadata %d, from disk %d",
 636                   be->bytenr, be->len, be->num_refs, be->metadata,
 637                   be->from_disk);
 638
 639         for (n = rb_first(&be->refs); n; n = rb_next(n)) {
 640                 ref = rb_entry(n, struct ref_entry, node);
 641                 btrfs_err(fs_info,
 642 "  ref root %llu, parent %llu, owner %llu, offset %llu, num_refs %llu",
 643                           ref->root_objectid, ref->parent, ref->owner,
 644                           ref->offset, ref->num_refs);
 645         }
 646
 647         for (n = rb_first(&be->roots); n; n = rb_next(n)) {
 648                 re = rb_entry(n, struct root_entry, node);
 649                 btrfs_err(fs_info, "  root entry %llu, num_refs %llu",
 650                           re->root_objectid, re->num_refs);
 651         }
 652
 653         list_for_each_entry(ra, &be->actions, list)
 654                 dump_ref_action(fs_info, ra);
 655 }
 656
 657 /*
 658  * Called when we modify a ref for a bytenr.
 659  *
 660  * This will add an action item to the given bytenr and do sanity checks to make
 661  * sure we haven't messed something up.  If we are making a new allocation and
 662  * this block entry has history we will delete all previous actions as long as
 663  * our sanity checks pass as they are no longer needed.
 664  */
 665 int btrfs_ref_tree_mod(struct btrfs_fs_info *fs_info,
 666                        struct btrfs_ref *generic_ref)
 667 {
 668         struct ref_entry *ref = NULL, *exist;
 669         struct ref_action *ra = NULL;
 670         struct block_entry *be = NULL;
 671         struct root_entry *re = NULL;
 672         int action = generic_ref->action;
 673         int ret = 0;
 674         bool metadata;
 675         u64 bytenr = generic_ref->bytenr;
 676         u64 num_bytes = generic_ref->len;
 677         u64 parent = generic_ref->parent;
 678         u64 ref_root = 0;
 679         u64 owner = 0;
 680         u64 offset = 0;
 681
 682         if (!btrfs_test_opt(fs_info, REF_VERIFY))
 683                 return 0;
 684
 685         if (generic_ref->type == BTRFS_REF_METADATA) {
 686                 if (!parent)
 687                         ref_root = generic_ref->tree_ref.ref_root;
 688                 owner = generic_ref->tree_ref.level;
 689         } else if (!parent) {
 690                 ref_root = generic_ref->data_ref.ref_root;
 691                 owner = generic_ref->data_ref.ino;
 692                 offset = generic_ref->data_ref.offset;
 693         }
 694         metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
 695
 696         ref = kzalloc(sizeof(struct ref_entry), GFP_NOFS);
 697         ra = kmalloc(sizeof(struct ref_action), GFP_NOFS);
 698         if (!ra || !ref) {
 699                 kfree(ref);
 700                 kfree(ra);
 701                 ret = -ENOMEM;
 702                 goto out;
 703         }
 704
 705         ref->parent = parent;
 706         ref->owner = owner;
 707         ref->root_objectid = ref_root;
 708         ref->offset = offset;
 709         ref->num_refs = (action == BTRFS_DROP_DELAYED_REF) ? -1 : 1;
 710
 711         memcpy(&ra->ref, ref, sizeof(struct ref_entry));
 712         /*
 713          * Save the extra info from the delayed ref in the ref action to make it
 714          * easier to figure out what is happening.  The real ref's we add to the
 715          * ref tree need to reflect what we save on disk so it matches any
 716          * on-disk refs we pre-loaded.
 717          */
 718         ra->ref.owner = owner;
 719         ra->ref.offset = offset;
 720         ra->ref.root_objectid = ref_root;
 721         __save_stack_trace(ra);
 722
 723         INIT_LIST_HEAD(&ra->list);
 724         ra->action = action;
 725         ra->root = generic_ref->real_root;
 726
 727         /*
 728          * This is an allocation, preallocate the block_entry in case we haven't
 729          * used it before.
 730          */
 731         ret = -EINVAL;
 732         if (action == BTRFS_ADD_DELAYED_EXTENT) {
 733                 /*
 734                  * For subvol_create we'll just pass in whatever the parent root
 735                  * is and the new root objectid, so let's not treat the passed
 736                  * in root as if it really has a ref for this bytenr.
 737                  */
 738                 be = add_block_entry(fs_info, bytenr, num_bytes, ref_root);
 739                 if (IS_ERR(be)) {
 740                         kfree(ref);
 741                         kfree(ra);
 742                         ret = PTR_ERR(be);
 743                         goto out;
 744                 }
 745                 be->num_refs++;
 746                 if (metadata)
 747                         be->metadata = 1;
 748
 749                 if (be->num_refs != 1) {
 750                         btrfs_err(fs_info,
 751                         "re-allocated a block that still has references to it!");
 752                         dump_block_entry(fs_info, be);
 753                         dump_ref_action(fs_info, ra);
 754                         kfree(ref);
 755                         kfree(ra);
 756                         goto out_unlock;
 757                 }
 758
 759                 while (!list_empty(&be->actions)) {
 760                         struct ref_action *tmp;
 761
 762                         tmp = list_first_entry(&be->actions, struct ref_action,
 763                                                list);
 764                         list_del(&tmp->list);
 765                         kfree(tmp);
 766                 }
 767         } else {
 768                 struct root_entry *tmp;
 769
 770                 if (!parent) {
 771                         re = kmalloc(sizeof(struct root_entry), GFP_NOFS);
 772                         if (!re) {
 773                                 kfree(ref);
 774                                 kfree(ra);
 775                                 ret = -ENOMEM;
 776                                 goto out;
 777                         }
 778                         /*
 779                          * This is the root that is modifying us, so it's the
 780                          * one we want to lookup below when we modify the
 781                          * re->num_refs.
 782                          */
 783                         ref_root = generic_ref->real_root;
 784                         re->root_objectid = generic_ref->real_root;
 785                         re->num_refs = 0;
 786                 }
 787
 788                 spin_lock(&fs_info->ref_verify_lock);
 789                 be = lookup_block_entry(&fs_info->block_tree, bytenr);
 790                 if (!be) {
 791                         btrfs_err(fs_info,
 792 "trying to do action %d to bytenr %llu num_bytes %llu but there is no existing entry!",
 793                                   action, bytenr, num_bytes);
 794                         dump_ref_action(fs_info, ra);
 795                         kfree(ref);
 796                         kfree(ra);
 797                         kfree(re);
 798                         goto out_unlock;
 799                 } else if (be->num_refs == 0) {
 800                         btrfs_err(fs_info,
 801                 "trying to do action %d for a bytenr that has 0 total references",
 802                                 action);
 803                         dump_block_entry(fs_info, be);
 804                         dump_ref_action(fs_info, ra);
 805                         kfree(ref);
 806                         kfree(ra);
 807                         kfree(re);
 808                         goto out_unlock;
 809                 }
 810
 811                 if (!parent) {
 812                         tmp = insert_root_entry(&be->roots, re);
 813                         if (tmp) {
 814                                 kfree(re);
 815                                 re = tmp;
 816                         }
 817                 }
 818         }
 819
 820         exist = insert_ref_entry(&be->refs, ref);
 821         if (exist) {
 822                 if (action == BTRFS_DROP_DELAYED_REF) {
 823                         if (exist->num_refs == 0) {
 824                                 btrfs_err(fs_info,
 825 "dropping a ref for a existing root that doesn't have a ref on the block");
 826                                 dump_block_entry(fs_info, be);
 827                                 dump_ref_action(fs_info, ra);
 828                                 kfree(ref);
 829                                 kfree(ra);
 830                                 goto out_unlock;
 831                         }
 832                         exist->num_refs--;
 833                         if (exist->num_refs == 0) {
 834                                 rb_erase(&exist->node, &be->refs);
 835                                 kfree(exist);
 836                         }
 837                 } else if (!be->metadata) {
 838                         exist->num_refs++;
 839                 } else {
 840                         btrfs_err(fs_info,
 841 "attempting to add another ref for an existing ref on a tree block");
 842                         dump_block_entry(fs_info, be);
 843                         dump_ref_action(fs_info, ra);
 844                         kfree(ref);
 845                         kfree(ra);
 846                         goto out_unlock;
 847                 }
 848                 kfree(ref);
 849         } else {
 850                 if (action == BTRFS_DROP_DELAYED_REF) {
 851                         btrfs_err(fs_info,
 852 "dropping a ref for a root that doesn't have a ref on the block");
 853                         dump_block_entry(fs_info, be);
 854                         dump_ref_action(fs_info, ra);
 855                         kfree(ref);
 856                         kfree(ra);
 857                         goto out_unlock;
 858                 }
 859         }
 860
 861         if (!parent && !re) {
 862                 re = lookup_root_entry(&be->roots, ref_root);
 863                 if (!re) {
 864                         /*
 865                          * This shouldn't happen because we will add our re
 866                          * above when we lookup the be with !parent, but just in
 867                          * case catch this case so we don't panic because I
 868                          * didn't think of some other corner case.
 869                          */
 870                         btrfs_err(fs_info, "failed to find root %llu for %llu",
 871                                   generic_ref->real_root, be->bytenr);
 872                         dump_block_entry(fs_info, be);
 873                         dump_ref_action(fs_info, ra);
 874                         kfree(ra);
 875                         goto out_unlock;
 876                 }
 877         }
 878         if (action == BTRFS_DROP_DELAYED_REF) {
 879                 if (re)
 880                         re->num_refs--;
 881                 be->num_refs--;
 882         } else if (action == BTRFS_ADD_DELAYED_REF) {
 883                 be->num_refs++;
 884                 if (re)
 885                         re->num_refs++;
 886         }
 887         list_add_tail(&ra->list, &be->actions);
 888         ret = 0;
 889 out_unlock:
 890         spin_unlock(&fs_info->ref_verify_lock);
 891 out:
 892         if (ret) {
 893                 btrfs_free_ref_cache(fs_info);
 894                 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
 895         }
 896         return ret;
 897 }
 898
 899 /* Free up the ref cache */
 900 void btrfs_free_ref_cache(struct btrfs_fs_info *fs_info)
 901 {
 902         struct block_entry *be;
 903         struct rb_node *n;
 904
 905         if (!btrfs_test_opt(fs_info, REF_VERIFY))
 906                 return;
 907
 908         spin_lock(&fs_info->ref_verify_lock);
 909         while ((n = rb_first(&fs_info->block_tree))) {
 910                 be = rb_entry(n, struct block_entry, node);
 911                 rb_erase(&be->node, &fs_info->block_tree);
 912                 free_block_entry(be);
 913                 cond_resched_lock(&fs_info->ref_verify_lock);
 914         }
 915         spin_unlock(&fs_info->ref_verify_lock);
 916 }
 917
 918 void btrfs_free_ref_tree_range(struct btrfs_fs_info *fs_info, u64 start,
 919                                u64 len)
 920 {
 921         struct block_entry *be = NULL, *entry;
 922         struct rb_node *n;
 923
 924         if (!btrfs_test_opt(fs_info, REF_VERIFY))
 925                 return;
 926
 927         spin_lock(&fs_info->ref_verify_lock);
 928         n = fs_info->block_tree.rb_node;
 929         while (n) {
 930                 entry = rb_entry(n, struct block_entry, node);
 931                 if (entry->bytenr < start) {
 932                         n = n->rb_right;
 933                 } else if (entry->bytenr > start) {
 934                         n = n->rb_left;
 935                 } else {
 936                         be = entry;
 937                         break;
 938                 }
 939                 /* We want to get as close to start as possible */
 940                 if (be == NULL ||
 941                     (entry->bytenr < start && be->bytenr > start) ||
 942                     (entry->bytenr < start && entry->bytenr > be->bytenr))
 943                         be = entry;
 944         }
 945
 946         /*
 947          * Could have an empty block group, maybe have something to check for
 948          * this case to verify we were actually empty?
 949          */
 950         if (!be) {
 951                 spin_unlock(&fs_info->ref_verify_lock);
 952                 return;
 953         }
 954
 955         n = &be->node;
 956         while (n) {
 957                 be = rb_entry(n, struct block_entry, node);
 958                 n = rb_next(n);
 959                 if (be->bytenr < start && be->bytenr + be->len > start) {
 960                         btrfs_err(fs_info,
 961                                 "block entry overlaps a block group [%llu,%llu]!",
 962                                 start, len);
 963                         dump_block_entry(fs_info, be);
 964                         continue;
 965                 }
 966                 if (be->bytenr < start)
 967                         continue;
 968                 if (be->bytenr >= start + len)
 969                         break;
 970                 if (be->bytenr + be->len > start + len) {
 971                         btrfs_err(fs_info,
 972                                 "block entry overlaps a block group [%llu,%llu]!",
 973                                 start, len);
 974                         dump_block_entry(fs_info, be);
 975                 }
 976                 rb_erase(&be->node, &fs_info->block_tree);
 977                 free_block_entry(be);
 978         }
 979         spin_unlock(&fs_info->ref_verify_lock);
 980 }
 981
 982 /* Walk down all roots and build the ref tree, meant to be called at mount */
 983 int btrfs_build_ref_tree(struct btrfs_fs_info *fs_info)
 984 {
 985         struct btrfs_root *extent_root;
 986         struct btrfs_path *path;
 987         struct extent_buffer *eb;
 988         int tree_block_level = 0;
 989         u64 bytenr = 0, num_bytes = 0;
 990         int ret, level;
 991
 992         if (!btrfs_test_opt(fs_info, REF_VERIFY))
 993                 return 0;
 994
 995         path = btrfs_alloc_path();
 996         if (!path)
 997                 return -ENOMEM;
 998
 999         extent_root = btrfs_extent_root(fs_info, 0);
1000         eb = btrfs_read_lock_root_node(extent_root);
1001         level = btrfs_header_level(eb);
1002         path->nodes[level] = eb;
1003         path->slots[level] = 0;
1004         path->locks[level] = BTRFS_READ_LOCK;
1005
1006         while (1) {
1007                 /*
1008                  * We have to keep track of the bytenr/num_bytes we last hit
1009                  * because we could have run out of space for an inline ref, and
1010                  * would have had to added a ref key item which may appear on a
1011                  * different leaf from the original extent item.
1012                  */
1013                 ret = walk_down_tree(extent_root, path, level,
1014                                      &bytenr, &num_bytes, &tree_block_level);
1015                 if (ret)
1016                         break;
1017                 ret = walk_up_tree(path, &level);
1018                 if (ret < 0)
1019                         break;
1020                 if (ret > 0) {
1021                         ret = 0;
1022                         break;
1023                 }
1024         }
1025         if (ret) {
1026                 btrfs_free_ref_cache(fs_info);
1027                 btrfs_clear_opt(fs_info->mount_opt, REF_VERIFY);
1028         }
1029         btrfs_free_path(path);
1030         return ret;
1031 }