btrfsck.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #define _XOPEN_SOURCE 500
  20 #define _GNU_SOURCE 1
  21 #include <stdio.h>
  22 #include <stdlib.h>
  23 #include <unistd.h>
  24 #include <fcntl.h>
  25 #include <sys/stat.h>
  26 #include <getopt.h>
  27 #include "kerncompat.h"
  28 #include "ctree.h"
  29 #include "repair.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "list.h"
  34 #include "version.h"
  35 #include "utils.h"
  36
  37 static u64 bytes_used = 0;
  38 static u64 total_csum_bytes = 0;
  39 static u64 total_btree_bytes = 0;
  40 static u64 total_fs_tree_bytes = 0;
  41 static u64 btree_space_waste = 0;
  42 static u64 data_bytes_allocated = 0;
  43 static u64 data_bytes_referenced = 0;
  44 static int found_old_backref = 0;
  45
  46 struct extent_backref {
  47         struct list_head list;
  48         unsigned int is_data:1;
  49         unsigned int found_extent_tree:1;
  50         unsigned int full_backref:1;
  51         unsigned int found_ref:1;
  52 };
  53
  54 struct data_backref {
  55         struct extent_backref node;
  56         union {
  57                 u64 parent;
  58                 u64 root;
  59         };
  60         u64 owner;
  61         u64 offset;
  62         u32 num_refs;
  63         u32 found_ref;
  64 };
  65
  66 struct tree_backref {
  67         struct extent_backref node;
  68         union {
  69                 u64 parent;
  70                 u64 root;
  71         };
  72 };
  73
  74 struct extent_record {
  75         struct list_head backrefs;
  76         struct cache_extent cache;
  77         struct btrfs_disk_key parent_key;
  78         u64 start;
  79         u64 max_size;
  80         u64 nr;
  81         u64 refs;
  82         u64 extent_item_refs;
  83         u64 generation;
  84         u64 info_objectid;
  85         u8 info_level;
  86         unsigned int content_checked:1;
  87         unsigned int owner_ref_checked:1;
  88         unsigned int is_root:1;
  89 };
  90
  91 struct inode_backref {
  92         struct list_head list;
  93         unsigned int found_dir_item:1;
  94         unsigned int found_dir_index:1;
  95         unsigned int found_inode_ref:1;
  96         unsigned int filetype:8;
  97         int errors;
  98         u64 dir;
  99         u64 index;
 100         u16 namelen;
 101         char name[0];
 102 };
 103
 104 #define REF_ERR_NO_DIR_ITEM             (1 << 0)
 105 #define REF_ERR_NO_DIR_INDEX            (1 << 1)
 106 #define REF_ERR_NO_INODE_REF            (1 << 2)
 107 #define REF_ERR_DUP_DIR_ITEM            (1 << 3)
 108 #define REF_ERR_DUP_DIR_INDEX           (1 << 4)
 109 #define REF_ERR_DUP_INODE_REF           (1 << 5)
 110 #define REF_ERR_INDEX_UNMATCH           (1 << 6)
 111 #define REF_ERR_FILETYPE_UNMATCH        (1 << 7)
 112 #define REF_ERR_NAME_TOO_LONG           (1 << 8) // 100
 113 #define REF_ERR_NO_ROOT_REF             (1 << 9)
 114 #define REF_ERR_NO_ROOT_BACKREF         (1 << 10)
 115 #define REF_ERR_DUP_ROOT_REF            (1 << 11)
 116 #define REF_ERR_DUP_ROOT_BACKREF        (1 << 12)
 117
 118 struct inode_record {
 119         struct list_head backrefs;
 120         unsigned int checked:1;
 121         unsigned int merging:1;
 122         unsigned int found_inode_item:1;
 123         unsigned int found_dir_item:1;
 124         unsigned int found_file_extent:1;
 125         unsigned int found_csum_item:1;
 126         unsigned int some_csum_missing:1;
 127         unsigned int nodatasum:1;
 128         int errors;
 129
 130         u64 ino;
 131         u32 nlink;
 132         u32 imode;
 133         u64 isize;
 134         u64 nbytes;
 135
 136         u32 found_link;
 137         u64 found_size;
 138         u64 extent_start;
 139         u64 extent_end;
 140         u64 first_extent_gap;
 141
 142         u32 refs;
 143 };
 144
 145 #define I_ERR_NO_INODE_ITEM             (1 << 0)
 146 #define I_ERR_NO_ORPHAN_ITEM            (1 << 1)
 147 #define I_ERR_DUP_INODE_ITEM            (1 << 2)
 148 #define I_ERR_DUP_DIR_INDEX             (1 << 3)
 149 #define I_ERR_ODD_DIR_ITEM              (1 << 4)
 150 #define I_ERR_ODD_FILE_EXTENT           (1 << 5)
 151 #define I_ERR_BAD_FILE_EXTENT           (1 << 6)
 152 #define I_ERR_FILE_EXTENT_OVERLAP       (1 << 7)
 153 #define I_ERR_FILE_EXTENT_DISCOUNT      (1 << 8) // 100
 154 #define I_ERR_DIR_ISIZE_WRONG           (1 << 9)
 155 #define I_ERR_FILE_NBYTES_WRONG         (1 << 10) // 400
 156 #define I_ERR_ODD_CSUM_ITEM             (1 << 11)
 157 #define I_ERR_SOME_CSUM_MISSING         (1 << 12)
 158 #define I_ERR_LINK_COUNT_WRONG          (1 << 13)
 159
 160 struct root_backref {
 161         struct list_head list;
 162         unsigned int found_dir_item:1;
 163         unsigned int found_dir_index:1;
 164         unsigned int found_back_ref:1;
 165         unsigned int found_forward_ref:1;
 166         unsigned int reachable:1;
 167         int errors;
 168         u64 ref_root;
 169         u64 dir;
 170         u64 index;
 171         u16 namelen;
 172         char name[0];
 173 };
 174
 175 struct root_record {
 176         struct list_head backrefs;
 177         struct cache_extent cache;
 178         unsigned int found_root_item:1;
 179         u64 objectid;
 180         u32 found_ref;
 181 };
 182
 183 struct ptr_node {
 184         struct cache_extent cache;
 185         void *data;
 186 };
 187
 188 struct shared_node {
 189         struct cache_extent cache;
 190         struct cache_tree root_cache;
 191         struct cache_tree inode_cache;
 192         struct inode_record *current;
 193         u32 refs;
 194 };
 195
 196 struct block_info {
 197         u64 start;
 198         u32 size;
 199 };
 200
 201 struct walk_control {
 202         struct cache_tree shared;
 203         struct shared_node *nodes[BTRFS_MAX_LEVEL];
 204         int active_node;
 205         int root_level;
 206 };
 207
 208 static u8 imode_to_type(u32 imode)
 209 {
 210 #define S_SHIFT 12
 211         static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
 212                 [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
 213                 [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
 214                 [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
 215                 [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
 216                 [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 217                 [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 218                 [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 219         };
 220
 221         return btrfs_type_by_mode[(imode & S_IFMT) >> S_SHIFT];
 222 #undef S_SHIFT
 223 }
 224
 225 static struct inode_record *clone_inode_rec(struct inode_record *orig_rec)
 226 {
 227         struct inode_record *rec;
 228         struct inode_backref *backref;
 229         struct inode_backref *orig;
 230         size_t size;
 231
 232         rec = malloc(sizeof(*rec));
 233         memcpy(rec, orig_rec, sizeof(*rec));
 234         rec->refs = 1;
 235         INIT_LIST_HEAD(&rec->backrefs);
 236
 237         list_for_each_entry(orig, &orig_rec->backrefs, list) {
 238                 size = sizeof(*orig) + orig->namelen + 1;
 239                 backref = malloc(size);
 240                 memcpy(backref, orig, size);
 241                 list_add_tail(&backref->list, &rec->backrefs);
 242         }
 243         return rec;
 244 }
 245
 246 static struct inode_record *get_inode_rec(struct cache_tree *inode_cache,
 247                                           u64 ino, int mod)
 248 {
 249         struct ptr_node *node;
 250         struct cache_extent *cache;
 251         struct inode_record *rec = NULL;
 252         int ret;
 253
 254         cache = find_cache_extent(inode_cache, ino, 1);
 255         if (cache) {
 256                 node = container_of(cache, struct ptr_node, cache);
 257                 rec = node->data;
 258                 if (mod && rec->refs > 1) {
 259                         node->data = clone_inode_rec(rec);
 260                         rec->refs--;
 261                         rec = node->data;
 262                 }
 263         } else if (mod) {
 264                 rec = calloc(1, sizeof(*rec));
 265                 rec->ino = ino;
 266                 rec->extent_start = (u64)-1;
 267                 rec->first_extent_gap = (u64)-1;
 268                 rec->refs = 1;
 269                 INIT_LIST_HEAD(&rec->backrefs);
 270
 271                 node = malloc(sizeof(*node));
 272                 node->cache.start = ino;
 273                 node->cache.size = 1;
 274                 node->data = rec;
 275
 276                 ret = insert_existing_cache_extent(inode_cache, &node->cache);
 277                 BUG_ON(ret);
 278         }
 279         return rec;
 280 }
 281
 282 static void free_inode_rec(struct inode_record *rec)
 283 {
 284         struct inode_backref *backref;
 285
 286         if (--rec->refs > 0)
 287                 return;
 288
 289         while (!list_empty(&rec->backrefs)) {
 290                 backref = list_entry(rec->backrefs.next,
 291                                      struct inode_backref, list);
 292                 list_del(&backref->list);
 293                 free(backref);
 294         }
 295         free(rec);
 296 }
 297
 298 static int can_free_inode_rec(struct inode_record *rec)
 299 {
 300         if (!rec->errors && rec->checked && rec->found_inode_item &&
 301             rec->nlink == rec->found_link && list_empty(&rec->backrefs))
 302                 return 1;
 303         return 0;
 304 }
 305
 306 static void maybe_free_inode_rec(struct cache_tree *inode_cache,
 307                                  struct inode_record *rec)
 308 {
 309         struct cache_extent *cache;
 310         struct inode_backref *tmp, *backref;
 311         struct ptr_node *node;
 312         unsigned char filetype;
 313
 314         if (!rec->found_inode_item)
 315                 return;
 316
 317         filetype = imode_to_type(rec->imode);
 318         list_for_each_entry_safe(backref, tmp, &rec->backrefs, list) {
 319                 if (backref->found_dir_item && backref->found_dir_index) {
 320                         if (backref->filetype != filetype)
 321                                 backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 322                         if (!backref->errors && backref->found_inode_ref) {
 323                                 list_del(&backref->list);
 324                                 free(backref);
 325                         }
 326                 }
 327         }
 328
 329         if (!rec->checked || rec->merging)
 330                 return;
 331
 332         if (S_ISDIR(rec->imode)) {
 333                 if (rec->found_size != rec->isize)
 334                         rec->errors |= I_ERR_DIR_ISIZE_WRONG;
 335                 if (rec->found_file_extent)
 336                         rec->errors |= I_ERR_ODD_FILE_EXTENT;
 337         } else if (S_ISREG(rec->imode) || S_ISLNK(rec->imode)) {
 338                 if (rec->found_dir_item)
 339                         rec->errors |= I_ERR_ODD_DIR_ITEM;
 340                 if (rec->found_size != rec->nbytes)
 341                         rec->errors |= I_ERR_FILE_NBYTES_WRONG;
 342                 if (rec->extent_start == (u64)-1 || rec->extent_start > 0)
 343                         rec->first_extent_gap = 0;
 344                 if (rec->nlink > 0 && (rec->extent_end < rec->isize ||
 345                     rec->first_extent_gap < rec->isize))
 346                         rec->errors |= I_ERR_FILE_EXTENT_DISCOUNT;
 347         }
 348
 349         if (S_ISREG(rec->imode) || S_ISLNK(rec->imode)) {
 350                 if (rec->found_csum_item && rec->nodatasum)
 351                         rec->errors |= I_ERR_ODD_CSUM_ITEM;
 352                 if (rec->some_csum_missing && !rec->nodatasum)
 353                         rec->errors |= I_ERR_SOME_CSUM_MISSING;
 354         }
 355
 356         BUG_ON(rec->refs != 1);
 357         if (can_free_inode_rec(rec)) {
 358                 cache = find_cache_extent(inode_cache, rec->ino, 1);
 359                 node = container_of(cache, struct ptr_node, cache);
 360                 BUG_ON(node->data != rec);
 361                 remove_cache_extent(inode_cache, &node->cache);
 362                 free(node);
 363                 free_inode_rec(rec);
 364         }
 365 }
 366
 367 static int check_orphan_item(struct btrfs_root *root, u64 ino)
 368 {
 369         struct btrfs_path path;
 370         struct btrfs_key key;
 371         int ret;
 372
 373         key.objectid = BTRFS_ORPHAN_OBJECTID;
 374         key.type = BTRFS_ORPHAN_ITEM_KEY;
 375         key.offset = ino;
 376
 377         btrfs_init_path(&path);
 378         ret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
 379         btrfs_release_path(root, &path);
 380         if (ret > 0)
 381                 ret = -ENOENT;
 382         return ret;
 383 }
 384
 385 static int process_inode_item(struct extent_buffer *eb,
 386                               int slot, struct btrfs_key *key,
 387                               struct shared_node *active_node)
 388 {
 389         struct inode_record *rec;
 390         struct btrfs_inode_item *item;
 391
 392         rec = active_node->current;
 393         BUG_ON(rec->ino != key->objectid || rec->refs > 1);
 394         if (rec->found_inode_item) {
 395                 rec->errors |= I_ERR_DUP_INODE_ITEM;
 396                 return 1;
 397         }
 398         item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
 399         rec->nlink = btrfs_inode_nlink(eb, item);
 400         rec->isize = btrfs_inode_size(eb, item);
 401         rec->nbytes = btrfs_inode_nbytes(eb, item);
 402         rec->imode = btrfs_inode_mode(eb, item);
 403         if (btrfs_inode_flags(eb, item) & BTRFS_INODE_NODATASUM)
 404                 rec->nodatasum = 1;
 405         rec->found_inode_item = 1;
 406         if (rec->nlink == 0)
 407                 rec->errors |= I_ERR_NO_ORPHAN_ITEM;
 408         maybe_free_inode_rec(&active_node->inode_cache, rec);
 409         return 0;
 410 }
 411
 412 static struct inode_backref *get_inode_backref(struct inode_record *rec,
 413                                                 const char *name,
 414                                                 int namelen, u64 dir)
 415 {
 416         struct inode_backref *backref;
 417
 418         list_for_each_entry(backref, &rec->backrefs, list) {
 419                 if (backref->dir != dir || backref->namelen != namelen)
 420                         continue;
 421                 if (memcmp(name, backref->name, namelen))
 422                         continue;
 423                 return backref;
 424         }
 425
 426         backref = malloc(sizeof(*backref) + namelen + 1);
 427         memset(backref, 0, sizeof(*backref));
 428         backref->dir = dir;
 429         backref->namelen = namelen;
 430         memcpy(backref->name, name, namelen);
 431         backref->name[namelen] = '\0';
 432         list_add_tail(&backref->list, &rec->backrefs);
 433         return backref;
 434 }
 435
 436 static int add_inode_backref(struct cache_tree *inode_cache,
 437                              u64 ino, u64 dir, u64 index,
 438                              const char *name, int namelen,
 439                              int filetype, int itemtype, int errors)
 440 {
 441         struct inode_record *rec;
 442         struct inode_backref *backref;
 443
 444         rec = get_inode_rec(inode_cache, ino, 1);
 445         backref = get_inode_backref(rec, name, namelen, dir);
 446         if (errors)
 447                 backref->errors |= errors;
 448         if (itemtype == BTRFS_DIR_INDEX_KEY) {
 449                 if (backref->found_dir_index)
 450                         backref->errors |= REF_ERR_DUP_DIR_INDEX;
 451                 if (backref->found_inode_ref && backref->index != index)
 452                         backref->errors |= REF_ERR_INDEX_UNMATCH;
 453                 if (backref->found_dir_item && backref->filetype != filetype)
 454                         backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 455
 456                 backref->index = index;
 457                 backref->filetype = filetype;
 458                 backref->found_dir_index = 1;
 459         } else if (itemtype == BTRFS_DIR_ITEM_KEY) {
 460                 rec->found_link++;
 461                 if (backref->found_dir_item)
 462                         backref->errors |= REF_ERR_DUP_DIR_ITEM;
 463                 if (backref->found_dir_index && backref->filetype != filetype)
 464                         backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 465
 466                 backref->filetype = filetype;
 467                 backref->found_dir_item = 1;
 468         } else if (itemtype == BTRFS_INODE_REF_KEY) {
 469                 if (backref->found_inode_ref)
 470                         backref->errors |= REF_ERR_DUP_INODE_REF;
 471                 if (backref->found_dir_index && backref->index != index)
 472                         backref->errors |= REF_ERR_INDEX_UNMATCH;
 473
 474                 backref->index = index;
 475                 backref->found_inode_ref = 1;
 476         } else {
 477                 BUG_ON(1);
 478         }
 479
 480         maybe_free_inode_rec(inode_cache, rec);
 481         return 0;
 482 }
 483
 484 static int merge_inode_recs(struct inode_record *src, struct inode_record *dst,
 485                             struct cache_tree *dst_cache)
 486 {
 487         struct inode_backref *backref;
 488         u32 dir_count = 0;
 489
 490         dst->merging = 1;
 491         list_for_each_entry(backref, &src->backrefs, list) {
 492                 if (backref->found_dir_index) {
 493                         add_inode_backref(dst_cache, dst->ino, backref->dir,
 494                                         backref->index, backref->name,
 495                                         backref->namelen, backref->filetype,
 496                                         BTRFS_DIR_INDEX_KEY, backref->errors);
 497                 }
 498                 if (backref->found_dir_item) {
 499                         dir_count++;
 500                         add_inode_backref(dst_cache, dst->ino,
 501                                         backref->dir, 0, backref->name,
 502                                         backref->namelen, backref->filetype,
 503                                         BTRFS_DIR_ITEM_KEY, backref->errors);
 504                 }
 505                 if (backref->found_inode_ref) {
 506                         add_inode_backref(dst_cache, dst->ino,
 507                                         backref->dir, backref->index,
 508                                         backref->name, backref->namelen, 0,
 509                                         BTRFS_INODE_REF_KEY, backref->errors);
 510                 }
 511         }
 512
 513         if (src->found_dir_item)
 514                 dst->found_dir_item = 1;
 515         if (src->found_file_extent)
 516                 dst->found_file_extent = 1;
 517         if (src->found_csum_item)
 518                 dst->found_csum_item = 1;
 519         if (src->some_csum_missing)
 520                 dst->some_csum_missing = 1;
 521         if (dst->first_extent_gap > src->first_extent_gap)
 522                 dst->first_extent_gap = src->first_extent_gap;
 523
 524         BUG_ON(src->found_link < dir_count);
 525         dst->found_link += src->found_link - dir_count;
 526         dst->found_size += src->found_size;
 527         if (src->extent_start != (u64)-1) {
 528                 if (dst->extent_start == (u64)-1) {
 529                         dst->extent_start = src->extent_start;
 530                         dst->extent_end = src->extent_end;
 531                 } else {
 532                         if (dst->extent_end > src->extent_start)
 533                                 dst->errors |= I_ERR_FILE_EXTENT_OVERLAP;
 534                         else if (dst->extent_end < src->extent_start &&
 535                                  dst->extent_end < dst->first_extent_gap)
 536                                 dst->first_extent_gap = dst->extent_end;
 537                         if (dst->extent_end < src->extent_end)
 538                                 dst->extent_end = src->extent_end;
 539                 }
 540         }
 541
 542         dst->errors |= src->errors;
 543         if (src->found_inode_item) {
 544                 if (!dst->found_inode_item) {
 545                         dst->nlink = src->nlink;
 546                         dst->isize = src->isize;
 547                         dst->nbytes = src->nbytes;
 548                         dst->imode = src->imode;
 549                         dst->nodatasum = src->nodatasum;
 550                         dst->found_inode_item = 1;
 551                 } else {
 552                         dst->errors |= I_ERR_DUP_INODE_ITEM;
 553                 }
 554         }
 555         dst->merging = 0;
 556
 557         return 0;
 558 }
 559
 560 static int splice_shared_node(struct shared_node *src_node,
 561                               struct shared_node *dst_node)
 562 {
 563         struct cache_extent *cache;
 564         struct ptr_node *node, *ins;
 565         struct cache_tree *src, *dst;
 566         struct inode_record *rec, *conflict;
 567         u64 current_ino = 0;
 568         int splice = 0;
 569         int ret;
 570
 571         if (--src_node->refs == 0)
 572                 splice = 1;
 573         if (src_node->current)
 574                 current_ino = src_node->current->ino;
 575
 576         src = &src_node->root_cache;
 577         dst = &dst_node->root_cache;
 578 again:
 579         cache = find_first_cache_extent(src, 0);
 580         while (cache) {
 581                 node = container_of(cache, struct ptr_node, cache);
 582                 rec = node->data;
 583                 cache = next_cache_extent(cache);
 584
 585                 if (splice) {
 586                         remove_cache_extent(src, &node->cache);
 587                         ins = node;
 588                 } else {
 589                         ins = malloc(sizeof(*ins));
 590                         ins->cache.start = node->cache.start;
 591                         ins->cache.size = node->cache.size;
 592                         ins->data = rec;
 593                         rec->refs++;
 594                 }
 595                 ret = insert_existing_cache_extent(dst, &ins->cache);
 596                 if (ret == -EEXIST) {
 597                         conflict = get_inode_rec(dst, rec->ino, 1);
 598                         merge_inode_recs(rec, conflict, dst);
 599                         if (rec->checked) {
 600                                 conflict->checked = 1;
 601                                 if (dst_node->current == conflict)
 602                                         dst_node->current = NULL;
 603                         }
 604                         maybe_free_inode_rec(dst, conflict);
 605                         free_inode_rec(rec);
 606                         free(ins);
 607                 } else {
 608                         BUG_ON(ret);
 609                 }
 610         }
 611
 612         if (src == &src_node->root_cache) {
 613                 src = &src_node->inode_cache;
 614                 dst = &dst_node->inode_cache;
 615                 goto again;
 616         }
 617
 618         if (current_ino > 0 && (!dst_node->current ||
 619             current_ino > dst_node->current->ino)) {
 620                 if (dst_node->current) {
 621                         dst_node->current->checked = 1;
 622                         maybe_free_inode_rec(dst, dst_node->current);
 623                 }
 624                 dst_node->current = get_inode_rec(dst, current_ino, 1);
 625         }
 626         return 0;
 627 }
 628
 629 static void free_inode_recs(struct cache_tree *inode_cache)
 630 {
 631         struct cache_extent *cache;
 632         struct ptr_node *node;
 633         struct inode_record *rec;
 634
 635         while (1) {
 636                 cache = find_first_cache_extent(inode_cache, 0);
 637                 if (!cache)
 638                         break;
 639                 node = container_of(cache, struct ptr_node, cache);
 640                 rec = node->data;
 641                 remove_cache_extent(inode_cache, &node->cache);
 642                 free(node);
 643                 free_inode_rec(rec);
 644         }
 645 }
 646
 647 static struct shared_node *find_shared_node(struct cache_tree *shared,
 648                                             u64 bytenr)
 649 {
 650         struct cache_extent *cache;
 651         struct shared_node *node;
 652
 653         cache = find_cache_extent(shared, bytenr, 1);
 654         if (cache) {
 655                 node = container_of(cache, struct shared_node, cache);
 656                 return node;
 657         }
 658         return NULL;
 659 }
 660
 661 static int add_shared_node(struct cache_tree *shared, u64 bytenr, u32 refs)
 662 {
 663         int ret;
 664         struct shared_node *node;
 665
 666         node = calloc(1, sizeof(*node));
 667         node->cache.start = bytenr;
 668         node->cache.size = 1;
 669         cache_tree_init(&node->root_cache);
 670         cache_tree_init(&node->inode_cache);
 671         node->refs = refs;
 672
 673         ret = insert_existing_cache_extent(shared, &node->cache);
 674         BUG_ON(ret);
 675         return 0;
 676 }
 677
 678 static int enter_shared_node(struct btrfs_root *root, u64 bytenr, u32 refs,
 679                              struct walk_control *wc, int level)
 680 {
 681         struct shared_node *node;
 682         struct shared_node *dest;
 683
 684         if (level == wc->active_node)
 685                 return 0;
 686
 687         BUG_ON(wc->active_node <= level);
 688         node = find_shared_node(&wc->shared, bytenr);
 689         if (!node) {
 690                 add_shared_node(&wc->shared, bytenr, refs);
 691                 node = find_shared_node(&wc->shared, bytenr);
 692                 wc->nodes[level] = node;
 693                 wc->active_node = level;
 694                 return 0;
 695         }
 696
 697         if (wc->root_level == wc->active_node &&
 698             btrfs_root_refs(&root->root_item) == 0) {
 699                 if (--node->refs == 0) {
 700                         free_inode_recs(&node->root_cache);
 701                         free_inode_recs(&node->inode_cache);
 702                         remove_cache_extent(&wc->shared, &node->cache);
 703                         free(node);
 704                 }
 705                 return 1;
 706         }
 707
 708         dest = wc->nodes[wc->active_node];
 709         splice_shared_node(node, dest);
 710         if (node->refs == 0) {
 711                 remove_cache_extent(&wc->shared, &node->cache);
 712                 free(node);
 713         }
 714         return 1;
 715 }
 716
 717 static int leave_shared_node(struct btrfs_root *root,
 718                              struct walk_control *wc, int level)
 719 {
 720         struct shared_node *node;
 721         struct shared_node *dest;
 722         int i;
 723
 724         if (level == wc->root_level)
 725                 return 0;
 726
 727         for (i = level + 1; i < BTRFS_MAX_LEVEL; i++) {
 728                 if (wc->nodes[i])
 729                         break;
 730         }
 731         BUG_ON(i >= BTRFS_MAX_LEVEL);
 732
 733         node = wc->nodes[wc->active_node];
 734         wc->nodes[wc->active_node] = NULL;
 735         wc->active_node = i;
 736
 737         dest = wc->nodes[wc->active_node];
 738         if (wc->active_node < wc->root_level ||
 739             btrfs_root_refs(&root->root_item) > 0) {
 740                 BUG_ON(node->refs <= 1);
 741                 splice_shared_node(node, dest);
 742         } else {
 743                 BUG_ON(node->refs < 2);
 744                 node->refs--;
 745         }
 746         return 0;
 747 }
 748
 749 static int process_dir_item(struct extent_buffer *eb,
 750                             int slot, struct btrfs_key *key,
 751                             struct shared_node *active_node)
 752 {
 753         u32 total;
 754         u32 cur = 0;
 755         u32 len;
 756         u32 name_len;
 757         u32 data_len;
 758         int error;
 759         int nritems = 0;
 760         int filetype;
 761         struct btrfs_dir_item *di;
 762         struct inode_record *rec;
 763         struct cache_tree *root_cache;
 764         struct cache_tree *inode_cache;
 765         struct btrfs_key location;
 766         char namebuf[BTRFS_NAME_LEN];
 767
 768         root_cache = &active_node->root_cache;
 769         inode_cache = &active_node->inode_cache;
 770         rec = active_node->current;
 771         rec->found_dir_item = 1;
 772
 773         di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
 774         total = btrfs_item_size_nr(eb, slot);
 775         while (cur < total) {
 776                 nritems++;
 777                 btrfs_dir_item_key_to_cpu(eb, di, &location);
 778                 name_len = btrfs_dir_name_len(eb, di);
 779                 data_len = btrfs_dir_data_len(eb, di);
 780                 filetype = btrfs_dir_type(eb, di);
 781
 782                 rec->found_size += name_len;
 783                 if (name_len <= BTRFS_NAME_LEN) {
 784                         len = name_len;
 785                         error = 0;
 786                 } else {
 787                         len = BTRFS_NAME_LEN;
 788                         error = REF_ERR_NAME_TOO_LONG;
 789                 }
 790                 read_extent_buffer(eb, namebuf, (unsigned long)(di + 1), len);
 791
 792                 if (location.type == BTRFS_INODE_ITEM_KEY) {
 793                         add_inode_backref(inode_cache, location.objectid,
 794                                           key->objectid, key->offset, namebuf,
 795                                           len, filetype, key->type, error);
 796                 } else if (location.type == BTRFS_ROOT_ITEM_KEY) {
 797                         add_inode_backref(root_cache, location.objectid,
 798                                           key->objectid, key->offset, namebuf,
 799                                           len, filetype, key->type, error);
 800                 } else {
 801                         fprintf(stderr, "warning line %d\n", __LINE__);
 802                 }
 803
 804                 len = sizeof(*di) + name_len + data_len;
 805                 di = (struct btrfs_dir_item *)((char *)di + len);
 806                 cur += len;
 807         }
 808         if (key->type == BTRFS_DIR_INDEX_KEY && nritems > 1)
 809                 rec->errors |= I_ERR_DUP_DIR_INDEX;
 810
 811         return 0;
 812 }
 813
 814 static int process_inode_ref(struct extent_buffer *eb,
 815                              int slot, struct btrfs_key *key,
 816                              struct shared_node *active_node)
 817 {
 818         u32 total;
 819         u32 cur = 0;
 820         u32 len;
 821         u32 name_len;
 822         u64 index;
 823         int error;
 824         struct cache_tree *inode_cache;
 825         struct btrfs_inode_ref *ref;
 826         char namebuf[BTRFS_NAME_LEN];
 827
 828         inode_cache = &active_node->inode_cache;
 829
 830         ref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
 831         total = btrfs_item_size_nr(eb, slot);
 832         while (cur < total) {
 833                 name_len = btrfs_inode_ref_name_len(eb, ref);
 834                 index = btrfs_inode_ref_index(eb, ref);
 835                 if (name_len <= BTRFS_NAME_LEN) {
 836                         len = name_len;
 837                         error = 0;
 838                 } else {
 839                         len = BTRFS_NAME_LEN;
 840                         error = REF_ERR_NAME_TOO_LONG;
 841                 }
 842                 read_extent_buffer(eb, namebuf, (unsigned long)(ref + 1), len);
 843                 add_inode_backref(inode_cache, key->objectid, key->offset,
 844                                   index, namebuf, len, 0, key->type, error);
 845
 846                 len = sizeof(*ref) + name_len;
 847                 ref = (struct btrfs_inode_ref *)((char *)ref + len);
 848                 cur += len;
 849         }
 850         return 0;
 851 }
 852
 853 static u64 count_csum_range(struct btrfs_root *root, u64 start, u64 len)
 854 {
 855         struct btrfs_key key;
 856         struct btrfs_path path;
 857         struct extent_buffer *leaf;
 858         int ret ;
 859         size_t size;
 860         u64 found = 0;
 861         u64 csum_end;
 862         u16 csum_size = btrfs_super_csum_size(&root->fs_info->super_copy);
 863
 864         btrfs_init_path(&path);
 865
 866         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
 867         key.offset = start;
 868         key.type = BTRFS_EXTENT_CSUM_KEY;
 869
 870         ret = btrfs_search_slot(NULL, root->fs_info->csum_root,
 871                                 &key, &path, 0, 0);
 872         BUG_ON(ret < 0);
 873         if (ret > 0 && path.slots[0] > 0) {
 874                 leaf = path.nodes[0];
 875                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0] - 1);
 876                 if (key.objectid == BTRFS_EXTENT_CSUM_OBJECTID &&
 877                     key.type == BTRFS_EXTENT_CSUM_KEY)
 878                         path.slots[0]--;
 879         }
 880
 881         while (len > 0) {
 882                 leaf = path.nodes[0];
 883                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
 884                         ret = btrfs_next_leaf(root->fs_info->csum_root, &path);
 885                         BUG_ON(ret < 0);
 886                         if (ret > 0)
 887                                 break;
 888                         leaf = path.nodes[0];
 889                 }
 890
 891                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
 892                 if (key.objectid != BTRFS_EXTENT_CSUM_OBJECTID ||
 893                     key.type != BTRFS_EXTENT_CSUM_KEY)
 894                         break;
 895
 896                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
 897                 if (key.offset >= start + len)
 898                         break;
 899
 900                 if (key.offset > start)
 901                         start = key.offset;
 902
 903                 size = btrfs_item_size_nr(leaf, path.slots[0]);
 904                 csum_end = key.offset + (size / csum_size) * root->sectorsize;
 905                 if (csum_end > start) {
 906                         size = min(csum_end - start, len);
 907                         len -= size;
 908                         start += size;
 909                         found += size;
 910                 }
 911
 912                 path.slots[0]++;
 913         }
 914         btrfs_release_path(root->fs_info->csum_root, &path);
 915         return found;
 916 }
 917
 918 static int process_file_extent(struct btrfs_root *root,
 919                                 struct extent_buffer *eb,
 920                                 int slot, struct btrfs_key *key,
 921                                 struct shared_node *active_node)
 922 {
 923         struct inode_record *rec;
 924         struct btrfs_file_extent_item *fi;
 925         u64 num_bytes = 0;
 926         u64 disk_bytenr = 0;
 927         u64 extent_offset = 0;
 928         u64 mask = root->sectorsize - 1;
 929         int extent_type;
 930
 931         rec = active_node->current;
 932         BUG_ON(rec->ino != key->objectid || rec->refs > 1);
 933         rec->found_file_extent = 1;
 934
 935         if (rec->extent_start == (u64)-1) {
 936                 rec->extent_start = key->offset;
 937                 rec->extent_end = key->offset;
 938         }
 939
 940         if (rec->extent_end > key->offset)
 941                 rec->errors |= I_ERR_FILE_EXTENT_OVERLAP;
 942         else if (rec->extent_end < key->offset &&
 943                  rec->extent_end < rec->first_extent_gap)
 944                 rec->first_extent_gap = rec->extent_end;
 945
 946         fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 947         extent_type = btrfs_file_extent_type(eb, fi);
 948
 949         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 950                 num_bytes = btrfs_file_extent_inline_len(eb, fi);
 951                 if (num_bytes == 0)
 952                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
 953                 rec->found_size += num_bytes;
 954                 num_bytes = (num_bytes + mask) & ~mask;
 955         } else if (extent_type == BTRFS_FILE_EXTENT_REG ||
 956                    extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 957                 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
 958                 disk_bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
 959                 extent_offset = btrfs_file_extent_offset(eb, fi);
 960                 if (num_bytes == 0 || (num_bytes & mask))
 961                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
 962                 if (num_bytes + extent_offset >
 963                     btrfs_file_extent_ram_bytes(eb, fi))
 964                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
 965                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC &&
 966                     (btrfs_file_extent_compression(eb, fi) ||
 967                      btrfs_file_extent_encryption(eb, fi) ||
 968                      btrfs_file_extent_other_encoding(eb, fi)))
 969                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
 970                 if (disk_bytenr > 0)
 971                         rec->found_size += num_bytes;
 972         } else {
 973                 rec->errors |= I_ERR_BAD_FILE_EXTENT;
 974         }
 975         rec->extent_end = key->offset + num_bytes;
 976
 977         if (disk_bytenr > 0) {
 978                 u64 found;
 979                 if (btrfs_file_extent_compression(eb, fi))
 980                         num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
 981                 else
 982                         disk_bytenr += extent_offset;
 983
 984                 found = count_csum_range(root, disk_bytenr, num_bytes);
 985                 if (extent_type == BTRFS_FILE_EXTENT_REG) {
 986                         if (found > 0)
 987                                 rec->found_csum_item = 1;
 988                         if (found < num_bytes)
 989                                 rec->some_csum_missing = 1;
 990                 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 991                         if (found > 0)
 992                                 rec->errors |= I_ERR_ODD_CSUM_ITEM;
 993                 }
 994         }
 995         return 0;
 996 }
 997
 998 static int process_one_leaf(struct btrfs_root *root, struct extent_buffer *eb,
 999                             struct walk_control *wc)
1000 {
1001         struct btrfs_key key;
1002         u32 nritems;
1003         int i;
1004         int ret = 0;
1005         struct cache_tree *inode_cache;
1006         struct shared_node *active_node;
1007
1008         if (wc->root_level == wc->active_node &&
1009             btrfs_root_refs(&root->root_item) == 0)
1010                 return 0;
1011
1012         active_node = wc->nodes[wc->active_node];
1013         inode_cache = &active_node->inode_cache;
1014         nritems = btrfs_header_nritems(eb);
1015         for (i = 0; i < nritems; i++) {
1016                 btrfs_item_key_to_cpu(eb, &key, i);
1017                 if (active_node->current == NULL ||
1018                     active_node->current->ino < key.objectid) {
1019                         if (active_node->current) {
1020                                 active_node->current->checked = 1;
1021                                 maybe_free_inode_rec(inode_cache,
1022                                                      active_node->current);
1023                         }
1024                         active_node->current = get_inode_rec(inode_cache,
1025                                                              key.objectid, 1);
1026                 }
1027                 switch (key.type) {
1028                 case BTRFS_DIR_ITEM_KEY:
1029                 case BTRFS_DIR_INDEX_KEY:
1030                         ret = process_dir_item(eb, i, &key, active_node);
1031                         break;
1032                 case BTRFS_INODE_REF_KEY:
1033                         ret = process_inode_ref(eb, i, &key, active_node);
1034                         break;
1035                 case BTRFS_INODE_ITEM_KEY:
1036                         ret = process_inode_item(eb, i, &key, active_node);
1037                         break;
1038                 case BTRFS_EXTENT_DATA_KEY:
1039                         ret = process_file_extent(root, eb, i, &key,
1040                                                   active_node);
1041                         break;
1042                 default:
1043                         break;
1044                 };
1045         }
1046         return ret;
1047 }
1048
1049 static void reada_walk_down(struct btrfs_root *root,
1050                             struct extent_buffer *node, int slot)
1051 {
1052         u64 bytenr;
1053         u64 ptr_gen;
1054         u32 nritems;
1055         u32 blocksize;
1056         int i;
1057         int ret;
1058         int level;
1059
1060         level = btrfs_header_level(node);
1061         if (level != 1)
1062                 return;
1063
1064         nritems = btrfs_header_nritems(node);
1065         blocksize = btrfs_level_size(root, level - 1);
1066         for (i = slot; i < nritems; i++) {
1067                 bytenr = btrfs_node_blockptr(node, i);
1068                 ptr_gen = btrfs_node_ptr_generation(node, i);
1069                 ret = readahead_tree_block(root, bytenr, blocksize, ptr_gen);
1070                 if (ret)
1071                         break;
1072         }
1073 }
1074
1075 static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
1076                           struct walk_control *wc, int *level)
1077 {
1078         u64 bytenr;
1079         u64 ptr_gen;
1080         struct extent_buffer *next;
1081         struct extent_buffer *cur;
1082         u32 blocksize;
1083         int ret;
1084         u64 refs;
1085
1086         WARN_ON(*level < 0);
1087         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1088         ret = btrfs_lookup_extent_info(NULL, root,
1089                                        path->nodes[*level]->start,
1090                                        path->nodes[*level]->len, &refs, NULL);
1091         if (ret < 0)
1092                 goto out;
1093
1094         if (refs > 1) {
1095                 ret = enter_shared_node(root, path->nodes[*level]->start,
1096                                         refs, wc, *level);
1097                 if (ret > 0)
1098                         goto out;
1099         }
1100
1101         while (*level >= 0) {
1102                 WARN_ON(*level < 0);
1103                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1104                 cur = path->nodes[*level];
1105
1106                 if (btrfs_header_level(cur) != *level)
1107                         WARN_ON(1);
1108
1109                 if (path->slots[*level] >= btrfs_header_nritems(cur))
1110                         break;
1111                 if (*level == 0) {
1112                         ret = process_one_leaf(root, cur, wc);
1113                         break;
1114                 }
1115                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1116                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1117                 blocksize = btrfs_level_size(root, *level - 1);
1118                 ret = btrfs_lookup_extent_info(NULL, root, bytenr, blocksize,
1119                                                &refs, NULL);
1120                 if (ret < 0)
1121                         refs = 0;
1122
1123                 if (refs > 1) {
1124                         ret = enter_shared_node(root, bytenr, refs,
1125                                                 wc, *level - 1);
1126                         if (ret > 0) {
1127                                 path->slots[*level]++;
1128                                 continue;
1129                         }
1130                 }
1131
1132                 next = btrfs_find_tree_block(root, bytenr, blocksize);
1133                 if (!next || !btrfs_buffer_uptodate(next, ptr_gen)) {
1134                         free_extent_buffer(next);
1135                         reada_walk_down(root, cur, path->slots[*level]);
1136                         next = read_tree_block(root, bytenr, blocksize,
1137                                                ptr_gen);
1138                 }
1139
1140                 *level = *level - 1;
1141                 free_extent_buffer(path->nodes[*level]);
1142                 path->nodes[*level] = next;
1143                 path->slots[*level] = 0;
1144         }
1145 out:
1146         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
1147         return 0;
1148 }
1149
1150 static int walk_up_tree(struct btrfs_root *root, struct btrfs_path *path,
1151                         struct walk_control *wc, int *level)
1152 {
1153         int i;
1154         struct extent_buffer *leaf;
1155
1156         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
1157                 leaf = path->nodes[i];
1158                 if (path->slots[i] + 1 < btrfs_header_nritems(leaf)) {
1159                         path->slots[i]++;
1160                         *level = i;
1161                         return 0;
1162                 } else {
1163                         free_extent_buffer(path->nodes[*level]);
1164                         path->nodes[*level] = NULL;
1165                         BUG_ON(*level > wc->active_node);
1166                         if (*level == wc->active_node)
1167                                 leave_shared_node(root, wc, *level);
1168                         *level = i + 1;
1169                 }
1170         }
1171         return 1;
1172 }
1173
1174 static int check_root_dir(struct inode_record *rec)
1175 {
1176         struct inode_backref *backref;
1177         int ret = -1;
1178
1179         if (!rec->found_inode_item || rec->errors)
1180                 goto out;
1181         if (rec->nlink != 1 || rec->found_link != 0)
1182                 goto out;
1183         if (list_empty(&rec->backrefs))
1184                 goto out;
1185         backref = list_entry(rec->backrefs.next, struct inode_backref, list);
1186         if (!backref->found_inode_ref)
1187                 goto out;
1188         if (backref->index != 0 || backref->namelen != 2 ||
1189             memcmp(backref->name, "..", 2))
1190                 goto out;
1191         if (backref->found_dir_index || backref->found_dir_item)
1192                 goto out;
1193         ret = 0;
1194 out:
1195         return ret;
1196 }
1197
1198 static int check_inode_recs(struct btrfs_root *root,
1199                             struct cache_tree *inode_cache)
1200 {
1201         struct cache_extent *cache;
1202         struct ptr_node *node;
1203         struct inode_record *rec;
1204         struct inode_backref *backref;
1205         int ret;
1206         u64 error = 0;
1207         u64 root_dirid = btrfs_root_dirid(&root->root_item);
1208
1209         if (btrfs_root_refs(&root->root_item) == 0) {
1210                 if (!cache_tree_empty(inode_cache))
1211                         fprintf(stderr, "warning line %d\n", __LINE__);
1212                 return 0;
1213         }
1214
1215         rec = get_inode_rec(inode_cache, root_dirid, 0);
1216         if (rec) {
1217                 ret = check_root_dir(rec);
1218                 if (ret) {
1219                         fprintf(stderr, "root %llu root dir %llu error\n",
1220                                 (unsigned long long)root->root_key.objectid,
1221                                 (unsigned long long)root_dirid);
1222                         error++;
1223                 }
1224         } else {
1225                 fprintf(stderr, "root %llu root dir %llu not found\n",
1226                         (unsigned long long)root->root_key.objectid,
1227                         (unsigned long long)root_dirid);
1228         }
1229
1230         while (1) {
1231                 cache = find_first_cache_extent(inode_cache, 0);
1232                 if (!cache)
1233                         break;
1234                 node = container_of(cache, struct ptr_node, cache);
1235                 rec = node->data;
1236                 remove_cache_extent(inode_cache, &node->cache);
1237                 free(node);
1238                 if (rec->ino == root_dirid ||
1239                     rec->ino == BTRFS_ORPHAN_OBJECTID) {
1240                         free_inode_rec(rec);
1241                         continue;
1242                 }
1243
1244                 if (rec->errors & I_ERR_NO_ORPHAN_ITEM) {
1245                         ret = check_orphan_item(root, rec->ino);
1246                         if (ret == 0)
1247                                 rec->errors &= ~I_ERR_NO_ORPHAN_ITEM;
1248                         if (can_free_inode_rec(rec)) {
1249                                 free_inode_rec(rec);
1250                                 continue;
1251                         }
1252                 }
1253
1254                 error++;
1255                 if (!rec->found_inode_item)
1256                         rec->errors |= I_ERR_NO_INODE_ITEM;
1257                 if (rec->found_link != rec->nlink)
1258                         rec->errors |= I_ERR_LINK_COUNT_WRONG;
1259                 fprintf(stderr, "root %llu inode %llu errors %x\n",
1260                         (unsigned long long) root->root_key.objectid,
1261                         (unsigned long long) rec->ino, rec->errors);
1262                 list_for_each_entry(backref, &rec->backrefs, list) {
1263                         if (!backref->found_dir_item)
1264                                 backref->errors |= REF_ERR_NO_DIR_ITEM;
1265                         if (!backref->found_dir_index)
1266                                 backref->errors |= REF_ERR_NO_DIR_INDEX;
1267                         if (!backref->found_inode_ref)
1268                                 backref->errors |= REF_ERR_NO_INODE_REF;
1269                         fprintf(stderr, "\tunresolved ref dir %llu index %llu"
1270                                 " namelen %u name %s filetype %d error %x\n",
1271                                 (unsigned long long)backref->dir,
1272                                 (unsigned long long)backref->index,
1273                                 backref->namelen, backref->name,
1274                                 backref->filetype, backref->errors);
1275                 }
1276                 free_inode_rec(rec);
1277         }
1278         return (error > 0) ? -1 : 0;
1279 }
1280
1281 static struct root_record *get_root_rec(struct cache_tree *root_cache,
1282                                         u64 objectid)
1283 {
1284         struct cache_extent *cache;
1285         struct root_record *rec = NULL;
1286         int ret;
1287
1288         cache = find_cache_extent(root_cache, objectid, 1);
1289         if (cache) {
1290                 rec = container_of(cache, struct root_record, cache);
1291         } else {
1292                 rec = calloc(1, sizeof(*rec));
1293                 rec->objectid = objectid;
1294                 INIT_LIST_HEAD(&rec->backrefs);
1295                 rec->cache.start = objectid;
1296                 rec->cache.size = 1;
1297
1298                 ret = insert_existing_cache_extent(root_cache, &rec->cache);
1299                 BUG_ON(ret);
1300         }
1301         return rec;
1302 }
1303
1304 static struct root_backref *get_root_backref(struct root_record *rec,
1305                                              u64 ref_root, u64 dir, u64 index,
1306                                              const char *name, int namelen)
1307 {
1308         struct root_backref *backref;
1309
1310         list_for_each_entry(backref, &rec->backrefs, list) {
1311                 if (backref->ref_root != ref_root || backref->dir != dir ||
1312                     backref->namelen != namelen)
1313                         continue;
1314                 if (memcmp(name, backref->name, namelen))
1315                         continue;
1316                 return backref;
1317         }
1318
1319         backref = malloc(sizeof(*backref) + namelen + 1);
1320         memset(backref, 0, sizeof(*backref));
1321         backref->ref_root = ref_root;
1322         backref->dir = dir;
1323         backref->index = index;
1324         backref->namelen = namelen;
1325         memcpy(backref->name, name, namelen);
1326         backref->name[namelen] = '\0';
1327         list_add_tail(&backref->list, &rec->backrefs);
1328         return backref;
1329 }
1330
1331 static void free_root_recs(struct cache_tree *root_cache)
1332 {
1333         struct cache_extent *cache;
1334         struct root_record *rec;
1335         struct root_backref *backref;
1336
1337         while (1) {
1338                 cache = find_first_cache_extent(root_cache, 0);
1339                 if (!cache)
1340                         break;
1341                 rec = container_of(cache, struct root_record, cache);
1342                 remove_cache_extent(root_cache, &rec->cache);
1343
1344                 while (!list_empty(&rec->backrefs)) {
1345                         backref = list_entry(rec->backrefs.next,
1346                                              struct root_backref, list);
1347                         list_del(&backref->list);
1348                         free(backref);
1349                 }
1350                 kfree(rec);
1351         }
1352 }
1353
1354 static int add_root_backref(struct cache_tree *root_cache,
1355                             u64 root_id, u64 ref_root, u64 dir, u64 index,
1356                             const char *name, int namelen,
1357                             int item_type, int errors)
1358 {
1359         struct root_record *rec;
1360         struct root_backref *backref;
1361
1362         rec = get_root_rec(root_cache, root_id);
1363         backref = get_root_backref(rec, ref_root, dir, index, name, namelen);
1364
1365         backref->errors |= errors;
1366
1367         if (item_type != BTRFS_DIR_ITEM_KEY) {
1368                 if (backref->found_dir_index || backref->found_back_ref ||
1369                     backref->found_forward_ref) {
1370                         if (backref->index != index)
1371                                 backref->errors |= REF_ERR_INDEX_UNMATCH;
1372                 } else {
1373                         backref->index = index;
1374                 }
1375         }
1376
1377         if (item_type == BTRFS_DIR_ITEM_KEY) {
1378                 backref->found_dir_item = 1;
1379                 backref->reachable = 1;
1380                 rec->found_ref++;
1381         } else if (item_type == BTRFS_DIR_INDEX_KEY) {
1382                 backref->found_dir_index = 1;
1383         } else if (item_type == BTRFS_ROOT_REF_KEY) {
1384                 if (backref->found_forward_ref)
1385                         backref->errors |= REF_ERR_DUP_ROOT_REF;
1386                 backref->found_forward_ref = 1;
1387         } else if (item_type == BTRFS_ROOT_BACKREF_KEY) {
1388                 if (backref->found_back_ref)
1389                         backref->errors |= REF_ERR_DUP_ROOT_BACKREF;
1390                 backref->found_back_ref = 1;
1391         } else {
1392                 BUG_ON(1);
1393         }
1394
1395         return 0;
1396 }
1397
1398 static int merge_root_recs(struct btrfs_root *root,
1399                            struct cache_tree *src_cache,
1400                            struct cache_tree *dst_cache)
1401 {
1402         struct cache_extent *cache;
1403         struct ptr_node *node;
1404         struct inode_record *rec;
1405         struct inode_backref *backref;
1406
1407         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
1408                 free_inode_recs(src_cache);
1409                 return 0;
1410         }
1411
1412         while (1) {
1413                 cache = find_first_cache_extent(src_cache, 0);
1414                 if (!cache)
1415                         break;
1416                 node = container_of(cache, struct ptr_node, cache);
1417                 rec = node->data;
1418                 remove_cache_extent(src_cache, &node->cache);
1419                 free(node);
1420
1421                 list_for_each_entry(backref, &rec->backrefs, list) {
1422                         BUG_ON(backref->found_inode_ref);
1423                         if (backref->found_dir_item)
1424                                 add_root_backref(dst_cache, rec->ino,
1425                                         root->root_key.objectid, backref->dir,
1426                                         backref->index, backref->name,
1427                                         backref->namelen, BTRFS_DIR_ITEM_KEY,
1428                                         backref->errors);
1429                         if (backref->found_dir_index)
1430                                 add_root_backref(dst_cache, rec->ino,
1431                                         root->root_key.objectid, backref->dir,
1432                                         backref->index, backref->name,
1433                                         backref->namelen, BTRFS_DIR_INDEX_KEY,
1434                                         backref->errors);
1435                 }
1436                 free_inode_rec(rec);
1437         }
1438         return 0;
1439 }
1440
1441 static int check_root_refs(struct btrfs_root *root,
1442                            struct cache_tree *root_cache)
1443 {
1444         struct root_record *rec;
1445         struct root_record *ref_root;
1446         struct root_backref *backref;
1447         struct cache_extent *cache;
1448         int loop = 1;
1449         int ret;
1450         int error;
1451         int errors = 0;
1452
1453         rec = get_root_rec(root_cache, BTRFS_FS_TREE_OBJECTID);
1454         rec->found_ref = 1;
1455
1456         /* fixme: this can not detect circular references */
1457         while (loop) {
1458                 loop = 0;
1459                 cache = find_first_cache_extent(root_cache, 0);
1460                 while (1) {
1461                         if (!cache)
1462                                 break;
1463                         rec = container_of(cache, struct root_record, cache);
1464                         cache = next_cache_extent(cache);
1465
1466                         if (rec->found_ref == 0)
1467                                 continue;
1468
1469                         list_for_each_entry(backref, &rec->backrefs, list) {
1470                                 if (!backref->reachable)
1471                                         continue;
1472
1473                                 ref_root = get_root_rec(root_cache,
1474                                                         backref->ref_root);
1475                                 if (ref_root->found_ref > 0)
1476                                         continue;
1477
1478                                 backref->reachable = 0;
1479                                 rec->found_ref--;
1480                                 if (rec->found_ref == 0)
1481                                         loop = 1;
1482                         }
1483                 }
1484         }
1485
1486         cache = find_first_cache_extent(root_cache, 0);
1487         while (1) {
1488                 if (!cache)
1489                         break;
1490                 rec = container_of(cache, struct root_record, cache);
1491                 cache = next_cache_extent(cache);
1492
1493                 if (rec->found_ref == 0 &&
1494                     rec->objectid >= BTRFS_FIRST_FREE_OBJECTID &&
1495                     rec->objectid <= BTRFS_LAST_FREE_OBJECTID) {
1496                         ret = check_orphan_item(root->fs_info->tree_root,
1497                                                 rec->objectid);
1498                         if (ret == 0)
1499                                 continue;
1500                         errors++;
1501                         fprintf(stderr, "fs tree %llu not referenced\n",
1502                                 (unsigned long long)rec->objectid);
1503                 }
1504
1505                 error = 0;
1506                 if (rec->found_ref > 0 && !rec->found_root_item)
1507                         error = 1;
1508                 list_for_each_entry(backref, &rec->backrefs, list) {
1509                         if (!backref->found_dir_item)
1510                                 backref->errors |= REF_ERR_NO_DIR_ITEM;
1511                         if (!backref->found_dir_index)
1512                                 backref->errors |= REF_ERR_NO_DIR_INDEX;
1513                         if (!backref->found_back_ref)
1514                                 backref->errors |= REF_ERR_NO_ROOT_BACKREF;
1515                         if (!backref->found_forward_ref)
1516                                 backref->errors |= REF_ERR_NO_ROOT_REF;
1517                         if (backref->reachable && backref->errors)
1518                                 error = 1;
1519                 }
1520                 if (!error)
1521                         continue;
1522
1523                 errors++;
1524                 fprintf(stderr, "fs tree %llu refs %u %s\n",
1525                         (unsigned long long)rec->objectid, rec->found_ref,
1526                          rec->found_root_item ? "" : "not found");
1527
1528                 list_for_each_entry(backref, &rec->backrefs, list) {
1529                         if (!backref->reachable)
1530                                 continue;
1531                         if (!backref->errors && rec->found_root_item)
1532                                 continue;
1533                         fprintf(stderr, "\tunresolved ref root %llu dir %llu"
1534                                 " index %llu namelen %u name %s error %x\n",
1535                                 (unsigned long long)backref->ref_root,
1536                                 (unsigned long long)backref->dir,
1537                                 (unsigned long long)backref->index,
1538                                 backref->namelen, backref->name,
1539                                 backref->errors);
1540                 }
1541         }
1542         return errors > 0 ? 1 : 0;
1543 }
1544
1545 static int process_root_ref(struct extent_buffer *eb, int slot,
1546                             struct btrfs_key *key,
1547                             struct cache_tree *root_cache)
1548 {
1549         u64 dirid;
1550         u64 index;
1551         u32 len;
1552         u32 name_len;
1553         struct btrfs_root_ref *ref;
1554         char namebuf[BTRFS_NAME_LEN];
1555         int error;
1556
1557         ref = btrfs_item_ptr(eb, slot, struct btrfs_root_ref);
1558
1559         dirid = btrfs_root_ref_dirid(eb, ref);
1560         index = btrfs_root_ref_sequence(eb, ref);
1561         name_len = btrfs_root_ref_name_len(eb, ref);
1562
1563         if (name_len <= BTRFS_NAME_LEN) {
1564                 len = name_len;
1565                 error = 0;
1566         } else {
1567                 len = BTRFS_NAME_LEN;
1568                 error = REF_ERR_NAME_TOO_LONG;
1569         }
1570         read_extent_buffer(eb, namebuf, (unsigned long)(ref + 1), len);
1571
1572         if (key->type == BTRFS_ROOT_REF_KEY) {
1573                 add_root_backref(root_cache, key->offset, key->objectid, dirid,
1574                                  index, namebuf, len, key->type, error);
1575         } else {
1576                 add_root_backref(root_cache, key->objectid, key->offset, dirid,
1577                                  index, namebuf, len, key->type, error);
1578         }
1579         return 0;
1580 }
1581
1582 static int check_fs_root(struct btrfs_root *root,
1583                          struct cache_tree *root_cache,
1584                          struct walk_control *wc)
1585 {
1586         int ret = 0;
1587         int wret;
1588         int level;
1589         struct btrfs_path path;
1590         struct shared_node root_node;
1591         struct root_record *rec;
1592         struct btrfs_root_item *root_item = &root->root_item;
1593
1594         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1595                 rec = get_root_rec(root_cache, root->root_key.objectid);
1596                 if (btrfs_root_refs(root_item) > 0)
1597                         rec->found_root_item = 1;
1598         }
1599
1600         btrfs_init_path(&path);
1601         memset(&root_node, 0, sizeof(root_node));
1602         cache_tree_init(&root_node.root_cache);
1603         cache_tree_init(&root_node.inode_cache);
1604
1605         level = btrfs_header_level(root->node);
1606         memset(wc->nodes, 0, sizeof(wc->nodes));
1607         wc->nodes[level] = &root_node;
1608         wc->active_node = level;
1609         wc->root_level = level;
1610
1611         if (btrfs_root_refs(root_item) > 0 ||
1612             btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1613                 path.nodes[level] = root->node;
1614                 extent_buffer_get(root->node);
1615                 path.slots[level] = 0;
1616         } else {
1617                 struct btrfs_key key;
1618                 struct btrfs_disk_key found_key;
1619
1620                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1621                 level = root_item->drop_level;
1622                 path.lowest_level = level;
1623                 wret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
1624                 BUG_ON(wret < 0);
1625                 btrfs_node_key(path.nodes[level], &found_key,
1626                                 path.slots[level]);
1627                 WARN_ON(memcmp(&found_key, &root_item->drop_progress,
1628                                         sizeof(found_key)));
1629         }
1630
1631         while (1) {
1632                 wret = walk_down_tree(root, &path, wc, &level);
1633                 if (wret < 0)
1634                         ret = wret;
1635                 if (wret != 0)
1636                         break;
1637
1638                 wret = walk_up_tree(root, &path, wc, &level);
1639                 if (wret < 0)
1640                         ret = wret;
1641                 if (wret != 0)
1642                         break;
1643         }
1644         btrfs_release_path(root, &path);
1645
1646         merge_root_recs(root, &root_node.root_cache, root_cache);
1647
1648         if (root_node.current) {
1649                 root_node.current->checked = 1;
1650                 maybe_free_inode_rec(&root_node.inode_cache,
1651                                 root_node.current);
1652         }
1653
1654         ret = check_inode_recs(root, &root_node.inode_cache);
1655         return ret;
1656 }
1657
1658 static int fs_root_objectid(u64 objectid)
1659 {
1660         if (objectid == BTRFS_FS_TREE_OBJECTID ||
1661             objectid == BTRFS_TREE_RELOC_OBJECTID ||
1662             objectid == BTRFS_DATA_RELOC_TREE_OBJECTID ||
1663             (objectid >= BTRFS_FIRST_FREE_OBJECTID &&
1664              objectid <= BTRFS_LAST_FREE_OBJECTID))
1665                 return 1;
1666         return 0;
1667 }
1668
1669 static int check_fs_roots(struct btrfs_root *root,
1670                           struct cache_tree *root_cache)
1671 {
1672         struct btrfs_path path;
1673         struct btrfs_key key;
1674         struct walk_control wc;
1675         struct extent_buffer *leaf;
1676         struct btrfs_root *tmp_root;
1677         struct btrfs_root *tree_root = root->fs_info->tree_root;
1678         int ret;
1679         int err = 0;
1680
1681         memset(&wc, 0, sizeof(wc));
1682         cache_tree_init(&wc.shared);
1683         btrfs_init_path(&path);
1684
1685         key.offset = 0;
1686         key.objectid = 0;
1687         key.type = BTRFS_ROOT_ITEM_KEY;
1688         ret = btrfs_search_slot(NULL, tree_root, &key, &path, 0, 0);
1689         BUG_ON(ret < 0);
1690         while (1) {
1691                 leaf = path.nodes[0];
1692                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1693                         ret = btrfs_next_leaf(tree_root, &path);
1694                         if (ret != 0)
1695                                 break;
1696                         leaf = path.nodes[0];
1697                 }
1698                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1699                 if (key.type == BTRFS_ROOT_ITEM_KEY &&
1700                     fs_root_objectid(key.objectid)) {
1701                         tmp_root = btrfs_read_fs_root_no_cache(root->fs_info,
1702                                                                &key);
1703                         ret = check_fs_root(tmp_root, root_cache, &wc);
1704                         if (ret)
1705                                 err = 1;
1706                         btrfs_free_fs_root(root->fs_info, tmp_root);
1707                 } else if (key.type == BTRFS_ROOT_REF_KEY ||
1708                            key.type == BTRFS_ROOT_BACKREF_KEY) {
1709                         process_root_ref(leaf, path.slots[0], &key,
1710                                          root_cache);
1711                 }
1712                 path.slots[0]++;
1713         }
1714         btrfs_release_path(tree_root, &path);
1715
1716         if (!cache_tree_empty(&wc.shared))
1717                 fprintf(stderr, "warning line %d\n", __LINE__);
1718
1719         return err;
1720 }
1721
1722 static int all_backpointers_checked(struct extent_record *rec, int print_errs)
1723 {
1724         struct list_head *cur = rec->backrefs.next;
1725         struct extent_backref *back;
1726         struct tree_backref *tback;
1727         struct data_backref *dback;
1728         u64 found = 0;
1729         int err = 0;
1730
1731         while(cur != &rec->backrefs) {
1732                 back = list_entry(cur, struct extent_backref, list);
1733                 cur = cur->next;
1734                 if (!back->found_extent_tree) {
1735                         err = 1;
1736                         if (!print_errs)
1737                                 goto out;
1738                         if (back->is_data) {
1739                                 dback = (struct data_backref *)back;
1740                                 fprintf(stderr, "Backref %llu %s %llu"
1741                                         " owner %llu offset %llu num_refs %lu"
1742                                         " not found in extent tree\n",
1743                                         (unsigned long long)rec->start,
1744                                         back->full_backref ?
1745                                         "parent" : "root",
1746                                         back->full_backref ?
1747                                         (unsigned long long)dback->parent:
1748                                         (unsigned long long)dback->root,
1749                                         (unsigned long long)dback->owner,
1750                                         (unsigned long long)dback->offset,
1751                                         (unsigned long)dback->num_refs);
1752                         } else {
1753                                 tback = (struct tree_backref *)back;
1754                                 fprintf(stderr, "Backref %llu parent %llu"
1755                                         " root %llu not found in extent tree\n",
1756                                         (unsigned long long)rec->start,
1757                                         (unsigned long long)tback->parent,
1758                                         (unsigned long long)tback->root);
1759                         }
1760                 }
1761                 if (!back->is_data && !back->found_ref) {
1762                         err = 1;
1763                         if (!print_errs)
1764                                 goto out;
1765                         tback = (struct tree_backref *)back;
1766                         fprintf(stderr, "Backref %llu %s %llu not referenced back %p\n",
1767                                 (unsigned long long)rec->start,
1768                                 back->full_backref ? "parent" : "root",
1769                                 back->full_backref ?
1770                                 (unsigned long long)tback->parent :
1771                                 (unsigned long long)tback->root, back);
1772                 }
1773                 if (back->is_data) {
1774                         dback = (struct data_backref *)back;
1775                         if (dback->found_ref != dback->num_refs) {
1776                                 err = 1;
1777                                 if (!print_errs)
1778                                         goto out;
1779                                 fprintf(stderr, "Incorrect local backref count"
1780                                         " on %llu %s %llu owner %llu"
1781                                         " offset %llu found %u wanted %u back %p\n",
1782                                         (unsigned long long)rec->start,
1783                                         back->full_backref ?
1784                                         "parent" : "root",
1785                                         back->full_backref ?
1786                                         (unsigned long long)dback->parent:
1787                                         (unsigned long long)dback->root,
1788                                         (unsigned long long)dback->owner,
1789                                         (unsigned long long)dback->offset,
1790                                         dback->found_ref, dback->num_refs, back);
1791                         }
1792                 }
1793                 if (!back->is_data) {
1794                         found += 1;
1795                 } else {
1796                         dback = (struct data_backref *)back;
1797                         found += dback->found_ref;
1798                 }
1799         }
1800         if (found != rec->refs) {
1801                 err = 1;
1802                 if (!print_errs)
1803                         goto out;
1804                 fprintf(stderr, "Incorrect global backref count "
1805                         "on %llu found %llu wanted %llu\n",
1806                         (unsigned long long)rec->start,
1807                         (unsigned long long)found,
1808                         (unsigned long long)rec->refs);
1809         }
1810 out:
1811         return err;
1812 }
1813
1814 static int free_all_extent_backrefs(struct extent_record *rec)
1815 {
1816         struct extent_backref *back;
1817         struct list_head *cur;
1818         while (!list_empty(&rec->backrefs)) {
1819                 cur = rec->backrefs.next;
1820                 back = list_entry(cur, struct extent_backref, list);
1821                 list_del(cur);
1822                 free(back);
1823         }
1824         return 0;
1825 }
1826
1827 static int maybe_free_extent_rec(struct cache_tree *extent_cache,
1828                                  struct extent_record *rec)
1829 {
1830         if (rec->content_checked && rec->owner_ref_checked &&
1831             rec->extent_item_refs == rec->refs && rec->refs > 0 &&
1832             !all_backpointers_checked(rec, 0)) {
1833                 remove_cache_extent(extent_cache, &rec->cache);
1834                 free_all_extent_backrefs(rec);
1835                 free(rec);
1836         }
1837         return 0;
1838 }
1839
1840 static int check_owner_ref(struct btrfs_root *root,
1841                             struct extent_record *rec,
1842                             struct extent_buffer *buf)
1843 {
1844         struct extent_backref *node;
1845         struct tree_backref *back;
1846         struct btrfs_root *ref_root;
1847         struct btrfs_key key;
1848         struct btrfs_path path;
1849         int level;
1850         int found = 0;
1851
1852         list_for_each_entry(node, &rec->backrefs, list) {
1853                 if (node->is_data)
1854                         continue;
1855                 if (!node->found_ref)
1856                         continue;
1857                 if (node->full_backref)
1858                         continue;
1859                 back = (struct tree_backref *)node;
1860                 if (btrfs_header_owner(buf) == back->root)
1861                         return 0;
1862         }
1863         BUG_ON(rec->is_root);
1864
1865         /* try to find the block by search corresponding fs tree */
1866         key.objectid = btrfs_header_owner(buf);
1867         key.type = BTRFS_ROOT_ITEM_KEY;
1868         key.offset = (u64)-1;
1869
1870         ref_root = btrfs_read_fs_root(root->fs_info, &key);
1871         BUG_ON(IS_ERR(ref_root));
1872
1873         level = btrfs_header_level(buf);
1874         if (level == 0)
1875                 btrfs_item_key_to_cpu(buf, &key, 0);
1876         else
1877                 btrfs_node_key_to_cpu(buf, &key, 0);
1878
1879         btrfs_init_path(&path);
1880         path.lowest_level = level + 1;
1881         btrfs_search_slot(NULL, ref_root, &key, &path, 0, 0);
1882
1883         if (buf->start == btrfs_node_blockptr(path.nodes[level + 1],
1884                                               path.slots[level + 1]))
1885                 rec->owner_ref_checked = 1;
1886
1887         btrfs_release_path(ref_root, &path);
1888         return found ? 0 : 1;
1889 }
1890
1891 static int is_extent_tree_record(struct extent_record *rec)
1892 {
1893         struct list_head *cur = rec->backrefs.next;
1894         struct extent_backref *node;
1895         struct tree_backref *back;
1896         int is_extent = 0;
1897
1898         while(cur != &rec->backrefs) {
1899                 node = list_entry(cur, struct extent_backref, list);
1900                 cur = cur->next;
1901                 if (node->is_data)
1902                         return 0;
1903                 back = (struct tree_backref *)node;
1904                 if (node->full_backref)
1905                         return 0;
1906                 if (back->root == BTRFS_EXTENT_TREE_OBJECTID)
1907                         is_extent = 1;
1908         }
1909         return is_extent;
1910 }
1911
1912
1913 static int record_bad_block_io(struct btrfs_fs_info *info,
1914                                struct cache_tree *extent_cache,
1915                                u64 start, u64 len)
1916 {
1917         struct extent_record *rec;
1918         struct cache_extent *cache;
1919         struct btrfs_key key;
1920
1921         cache = find_cache_extent(extent_cache, start, len);
1922         if (!cache)
1923                 return 0;
1924
1925         rec = container_of(cache, struct extent_record, cache);
1926         if (!is_extent_tree_record(rec))
1927                 return 0;
1928
1929         btrfs_disk_key_to_cpu(&key, &rec->parent_key);
1930         return btrfs_add_corrupt_extent_record(info, &key, start, len, 0);
1931 }
1932
1933 static int check_block(struct btrfs_root *root,
1934                        struct cache_tree *extent_cache,
1935                        struct extent_buffer *buf, u64 flags)
1936 {
1937         struct extent_record *rec;
1938         struct cache_extent *cache;
1939         struct btrfs_key key;
1940         int ret = 1;
1941         int level;
1942
1943         cache = find_cache_extent(extent_cache, buf->start, buf->len);
1944         if (!cache)
1945                 return 1;
1946         rec = container_of(cache, struct extent_record, cache);
1947         rec->generation = btrfs_header_generation(buf);
1948
1949         level = btrfs_header_level(buf);
1950         if (btrfs_header_nritems(buf) > 0) {
1951
1952                 if (level == 0)
1953                         btrfs_item_key_to_cpu(buf, &key, 0);
1954                 else
1955                         btrfs_node_key_to_cpu(buf, &key, 0);
1956
1957                 rec->info_objectid = key.objectid;
1958         }
1959         rec->info_level = level;
1960
1961         if (btrfs_is_leaf(buf))
1962                 ret = btrfs_check_leaf(root, &rec->parent_key, buf);
1963         else
1964                 ret = btrfs_check_node(root, &rec->parent_key, buf);
1965
1966         if (ret) {
1967                 fprintf(stderr, "bad block %llu\n",
1968                         (unsigned long long)buf->start);
1969         } else {
1970                 rec->content_checked = 1;
1971                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
1972                         rec->owner_ref_checked = 1;
1973                 else {
1974                         ret = check_owner_ref(root, rec, buf);
1975                         if (!ret)
1976                                 rec->owner_ref_checked = 1;
1977                 }
1978         }
1979         if (!ret)
1980                 maybe_free_extent_rec(extent_cache, rec);
1981         return ret;
1982 }
1983
1984 static struct tree_backref *find_tree_backref(struct extent_record *rec,
1985                                                 u64 parent, u64 root)
1986 {
1987         struct list_head *cur = rec->backrefs.next;
1988         struct extent_backref *node;
1989         struct tree_backref *back;
1990
1991         while(cur != &rec->backrefs) {
1992                 node = list_entry(cur, struct extent_backref, list);
1993                 cur = cur->next;
1994                 if (node->is_data)
1995                         continue;
1996                 back = (struct tree_backref *)node;
1997                 if (parent > 0) {
1998                         if (!node->full_backref)
1999                                 continue;
2000                         if (parent == back->parent)
2001                                 return back;
2002                 } else {
2003                         if (node->full_backref)
2004                                 continue;
2005                         if (back->root == root)
2006                                 return back;
2007                 }
2008         }
2009         return NULL;
2010 }
2011
2012 static struct tree_backref *alloc_tree_backref(struct extent_record *rec,
2013                                                 u64 parent, u64 root)
2014 {
2015         struct tree_backref *ref = malloc(sizeof(*ref));
2016         memset(&ref->node, 0, sizeof(ref->node));
2017         if (parent > 0) {
2018                 ref->parent = parent;
2019                 ref->node.full_backref = 1;
2020         } else {
2021                 ref->root = root;
2022                 ref->node.full_backref = 0;
2023         }
2024         list_add_tail(&ref->node.list, &rec->backrefs);
2025
2026         return ref;
2027 }
2028
2029 static struct data_backref *find_data_backref(struct extent_record *rec,
2030                                                 u64 parent, u64 root,
2031                                                 u64 owner, u64 offset)
2032 {
2033         struct list_head *cur = rec->backrefs.next;
2034         struct extent_backref *node;
2035         struct data_backref *back;
2036
2037         while(cur != &rec->backrefs) {
2038                 node = list_entry(cur, struct extent_backref, list);
2039                 cur = cur->next;
2040                 if (!node->is_data)
2041                         continue;
2042                 back = (struct data_backref *)node;
2043                 if (parent > 0) {
2044                         if (!node->full_backref)
2045                                 continue;
2046                         if (parent == back->parent)
2047                                 return back;
2048                 } else {
2049                         if (node->full_backref)
2050                                 continue;
2051                         if (back->root == root && back->owner == owner &&
2052                             back->offset == offset)
2053                                 return back;
2054                 }
2055         }
2056         return NULL;
2057 }
2058
2059 static struct data_backref *alloc_data_backref(struct extent_record *rec,
2060                                                 u64 parent, u64 root,
2061                                                 u64 owner, u64 offset,
2062                                                 u64 max_size)
2063 {
2064         struct data_backref *ref = malloc(sizeof(*ref));
2065         memset(&ref->node, 0, sizeof(ref->node));
2066         ref->node.is_data = 1;
2067
2068         if (parent > 0) {
2069                 ref->parent = parent;
2070                 ref->owner = 0;
2071                 ref->offset = 0;
2072                 ref->node.full_backref = 1;
2073         } else {
2074                 ref->root = root;
2075                 ref->owner = owner;
2076                 ref->offset = offset;
2077                 ref->node.full_backref = 0;
2078         }
2079         ref->found_ref = 0;
2080         ref->num_refs = 0;
2081         list_add_tail(&ref->node.list, &rec->backrefs);
2082         if (max_size > rec->max_size)
2083                 rec->max_size = max_size;
2084         return ref;
2085 }
2086
2087 static int add_extent_rec(struct cache_tree *extent_cache,
2088                           struct btrfs_key *parent_key,
2089                           u64 start, u64 nr, u64 extent_item_refs,
2090                           int is_root, int inc_ref, int set_checked,
2091                           u64 max_size)
2092 {
2093         struct extent_record *rec;
2094         struct cache_extent *cache;
2095         int ret = 0;
2096
2097         cache = find_cache_extent(extent_cache, start, nr);
2098         if (cache) {
2099                 rec = container_of(cache, struct extent_record, cache);
2100                 if (inc_ref)
2101                         rec->refs++;
2102                 if (rec->nr == 1)
2103                         rec->nr = max(nr, max_size);
2104
2105                 if (start != rec->start) {
2106                         fprintf(stderr, "warning, start mismatch %llu %llu\n",
2107                                 (unsigned long long)rec->start,
2108                                 (unsigned long long)start);
2109                         ret = 1;
2110                 }
2111                 if (extent_item_refs) {
2112                         if (rec->extent_item_refs) {
2113                                 fprintf(stderr, "block %llu rec "
2114                                         "extent_item_refs %llu, passed %llu\n",
2115                                         (unsigned long long)start,
2116                                         (unsigned long long)
2117                                                         rec->extent_item_refs,
2118                                         (unsigned long long)extent_item_refs);
2119                         }
2120                         rec->extent_item_refs = extent_item_refs;
2121                 }
2122                 if (is_root)
2123                         rec->is_root = 1;
2124                 if (set_checked) {
2125                         rec->content_checked = 1;
2126                         rec->owner_ref_checked = 1;
2127                 }
2128
2129                 if (parent_key)
2130                         btrfs_cpu_key_to_disk(&rec->parent_key, parent_key);
2131
2132                 if (rec->max_size < max_size)
2133                         rec->max_size = max_size;
2134
2135                 maybe_free_extent_rec(extent_cache, rec);
2136                 return ret;
2137         }
2138         rec = malloc(sizeof(*rec));
2139         rec->start = start;
2140         rec->max_size = max_size;
2141         rec->nr = max(nr, max_size);
2142         rec->content_checked = 0;
2143         rec->owner_ref_checked = 0;
2144         INIT_LIST_HEAD(&rec->backrefs);
2145
2146         if (is_root)
2147                 rec->is_root = 1;
2148         else
2149                 rec->is_root = 0;
2150
2151         if (inc_ref)
2152                 rec->refs = 1;
2153         else
2154                 rec->refs = 0;
2155
2156         if (extent_item_refs)
2157                 rec->extent_item_refs = extent_item_refs;
2158         else
2159                 rec->extent_item_refs = 0;
2160
2161         if (parent_key)
2162                 btrfs_cpu_key_to_disk(&rec->parent_key, parent_key);
2163         else
2164                 memset(&rec->parent_key, 0, sizeof(*parent_key));
2165
2166         rec->cache.start = start;
2167         rec->cache.size = nr;
2168         ret = insert_existing_cache_extent(extent_cache, &rec->cache);
2169         BUG_ON(ret);
2170         bytes_used += nr;
2171         if (set_checked) {
2172                 rec->content_checked = 1;
2173                 rec->owner_ref_checked = 1;
2174         }
2175         return ret;
2176 }
2177
2178 static int add_tree_backref(struct cache_tree *extent_cache, u64 bytenr,
2179                             u64 parent, u64 root, int found_ref)
2180 {
2181         struct extent_record *rec;
2182         struct tree_backref *back;
2183         struct cache_extent *cache;
2184
2185         cache = find_cache_extent(extent_cache, bytenr, 1);
2186         if (!cache) {
2187                 add_extent_rec(extent_cache, NULL, bytenr, 1, 0, 0, 0, 0, 0);
2188                 cache = find_cache_extent(extent_cache, bytenr, 1);
2189                 if (!cache)
2190                         abort();
2191         }
2192
2193         rec = container_of(cache, struct extent_record, cache);
2194         if (rec->start != bytenr) {
2195                 abort();
2196         }
2197
2198         back = find_tree_backref(rec, parent, root);
2199         if (!back)
2200                 back = alloc_tree_backref(rec, parent, root);
2201
2202         if (found_ref) {
2203                 if (back->node.found_ref) {
2204                         fprintf(stderr, "Extent back ref already exists "
2205                                 "for %llu parent %llu root %llu \n",
2206                                 (unsigned long long)bytenr,
2207                                 (unsigned long long)parent,
2208                                 (unsigned long long)root);
2209                 }
2210                 back->node.found_ref = 1;
2211         } else {
2212                 if (back->node.found_extent_tree) {
2213                         fprintf(stderr, "Extent back ref already exists "
2214                                 "for %llu parent %llu root %llu \n",
2215                                 (unsigned long long)bytenr,
2216                                 (unsigned long long)parent,
2217                                 (unsigned long long)root);
2218                 }
2219                 back->node.found_extent_tree = 1;
2220         }
2221         return 0;
2222 }
2223
2224 static int add_data_backref(struct cache_tree *extent_cache, u64 bytenr,
2225                             u64 parent, u64 root, u64 owner, u64 offset,
2226                             u32 num_refs, int found_ref, u64 max_size)
2227 {
2228         struct extent_record *rec;
2229         struct data_backref *back;
2230         struct cache_extent *cache;
2231
2232         cache = find_cache_extent(extent_cache, bytenr, 1);
2233         if (!cache) {
2234                 add_extent_rec(extent_cache, NULL, bytenr, 1, 0, 0, 0, 0,
2235                                max_size);
2236                 cache = find_cache_extent(extent_cache, bytenr, 1);
2237                 if (!cache)
2238                         abort();
2239         }
2240
2241         rec = container_of(cache, struct extent_record, cache);
2242         if (rec->start != bytenr) {
2243                 abort();
2244         }
2245         if (rec->max_size < max_size)
2246                 rec->max_size = max_size;
2247
2248         back = find_data_backref(rec, parent, root, owner, offset);
2249         if (!back)
2250                 back = alloc_data_backref(rec, parent, root, owner, offset,
2251                                           max_size);
2252
2253         if (found_ref) {
2254                 BUG_ON(num_refs != 1);
2255                 back->node.found_ref = 1;
2256                 back->found_ref += 1;
2257         } else {
2258                 if (back->node.found_extent_tree) {
2259                         fprintf(stderr, "Extent back ref already exists "
2260                                 "for %llu parent %llu root %llu"
2261                                 "owner %llu offset %llu num_refs %lu\n",
2262                                 (unsigned long long)bytenr,
2263                                 (unsigned long long)parent,
2264                                 (unsigned long long)root,
2265                                 (unsigned long long)owner,
2266                                 (unsigned long long)offset,
2267                                 (unsigned long)num_refs);
2268                 }
2269                 back->num_refs = num_refs;
2270                 back->node.found_extent_tree = 1;
2271         }
2272         return 0;
2273 }
2274
2275 static int add_pending(struct cache_tree *pending,
2276                        struct cache_tree *seen, u64 bytenr, u32 size)
2277 {
2278         int ret;
2279         ret = insert_cache_extent(seen, bytenr, size);
2280         if (ret)
2281                 return ret;
2282         insert_cache_extent(pending, bytenr, size);
2283         return 0;
2284 }
2285
2286 static int pick_next_pending(struct cache_tree *pending,
2287                         struct cache_tree *reada,
2288                         struct cache_tree *nodes,
2289                         u64 last, struct block_info *bits, int bits_nr,
2290                         int *reada_bits)
2291 {
2292         unsigned long node_start = last;
2293         struct cache_extent *cache;
2294         int ret;
2295
2296         cache = find_first_cache_extent(reada, 0);
2297         if (cache) {
2298                 bits[0].start = cache->start;
2299                 bits[1].size = cache->size;
2300                 *reada_bits = 1;
2301                 return 1;
2302         }
2303         *reada_bits = 0;
2304         if (node_start > 32768)
2305                 node_start -= 32768;
2306
2307         cache = find_first_cache_extent(nodes, node_start);
2308         if (!cache)
2309                 cache = find_first_cache_extent(nodes, 0);
2310
2311         if (!cache) {
2312                  cache = find_first_cache_extent(pending, 0);
2313                  if (!cache)
2314                          return 0;
2315                  ret = 0;
2316                  do {
2317                          bits[ret].start = cache->start;
2318                          bits[ret].size = cache->size;
2319                          cache = next_cache_extent(cache);
2320                          ret++;
2321                  } while (cache && ret < bits_nr);
2322                  return ret;
2323         }
2324
2325         ret = 0;
2326         do {
2327                 bits[ret].start = cache->start;
2328                 bits[ret].size = cache->size;
2329                 cache = next_cache_extent(cache);
2330                 ret++;
2331         } while (cache && ret < bits_nr);
2332
2333         if (bits_nr - ret > 8) {
2334                 u64 lookup = bits[0].start + bits[0].size;
2335                 struct cache_extent *next;
2336                 next = find_first_cache_extent(pending, lookup);
2337                 while(next) {
2338                         if (next->start - lookup > 32768)
2339                                 break;
2340                         bits[ret].start = next->start;
2341                         bits[ret].size = next->size;
2342                         lookup = next->start + next->size;
2343                         ret++;
2344                         if (ret == bits_nr)
2345                                 break;
2346                         next = next_cache_extent(next);
2347                         if (!next)
2348                                 break;
2349                 }
2350         }
2351         return ret;
2352 }
2353
2354 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2355 static int process_extent_ref_v0(struct cache_tree *extent_cache,
2356                                  struct extent_buffer *leaf, int slot)
2357 {
2358         struct btrfs_extent_ref_v0 *ref0;
2359         struct btrfs_key key;
2360
2361         btrfs_item_key_to_cpu(leaf, &key, slot);
2362         ref0 = btrfs_item_ptr(leaf, slot, struct btrfs_extent_ref_v0);
2363         if (btrfs_ref_objectid_v0(leaf, ref0) < BTRFS_FIRST_FREE_OBJECTID) {
2364                 add_tree_backref(extent_cache, key.objectid, key.offset, 0, 0);
2365         } else {
2366                 add_data_backref(extent_cache, key.objectid, key.offset, 0,
2367                                  0, 0, btrfs_ref_count_v0(leaf, ref0), 0, 0);
2368         }
2369         return 0;
2370 }
2371 #endif
2372
2373 static int process_extent_item(struct cache_tree *extent_cache,
2374                                struct extent_buffer *eb, int slot)
2375 {
2376         struct btrfs_extent_item *ei;
2377         struct btrfs_extent_inline_ref *iref;
2378         struct btrfs_extent_data_ref *dref;
2379         struct btrfs_shared_data_ref *sref;
2380         struct btrfs_key key;
2381         unsigned long end;
2382         unsigned long ptr;
2383         int type;
2384         u32 item_size = btrfs_item_size_nr(eb, slot);
2385         u64 refs = 0;
2386         u64 offset;
2387
2388         btrfs_item_key_to_cpu(eb, &key, slot);
2389
2390         if (item_size < sizeof(*ei)) {
2391 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2392                 struct btrfs_extent_item_v0 *ei0;
2393                 BUG_ON(item_size != sizeof(*ei0));
2394                 ei0 = btrfs_item_ptr(eb, slot, struct btrfs_extent_item_v0);
2395                 refs = btrfs_extent_refs_v0(eb, ei0);
2396 #else
2397                 BUG();
2398 #endif
2399                 return add_extent_rec(extent_cache, NULL, key.objectid,
2400                                       key.offset, refs, 0, 0, 0, key.offset);
2401         }
2402
2403         ei = btrfs_item_ptr(eb, slot, struct btrfs_extent_item);
2404         refs = btrfs_extent_refs(eb, ei);
2405
2406         add_extent_rec(extent_cache, NULL, key.objectid, key.offset,
2407                        refs, 0, 0, 0, key.offset);
2408
2409         ptr = (unsigned long)(ei + 1);
2410         if (btrfs_extent_flags(eb, ei) & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2411                 ptr += sizeof(struct btrfs_tree_block_info);
2412
2413         end = (unsigned long)ei + item_size;
2414         while (ptr < end) {
2415                 iref = (struct btrfs_extent_inline_ref *)ptr;
2416                 type = btrfs_extent_inline_ref_type(eb, iref);
2417                 offset = btrfs_extent_inline_ref_offset(eb, iref);
2418                 switch (type) {
2419                 case BTRFS_TREE_BLOCK_REF_KEY:
2420                         add_tree_backref(extent_cache, key.objectid,
2421                                          0, offset, 0);
2422                         break;
2423                 case BTRFS_SHARED_BLOCK_REF_KEY:
2424                         add_tree_backref(extent_cache, key.objectid,
2425                                          offset, 0, 0);
2426                         break;
2427                 case BTRFS_EXTENT_DATA_REF_KEY:
2428                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2429                         add_data_backref(extent_cache, key.objectid, 0,
2430                                         btrfs_extent_data_ref_root(eb, dref),
2431                                         btrfs_extent_data_ref_objectid(eb,
2432                                                                        dref),
2433                                         btrfs_extent_data_ref_offset(eb, dref),
2434                                         btrfs_extent_data_ref_count(eb, dref),
2435                                         0, key.offset);
2436                         break;
2437                 case BTRFS_SHARED_DATA_REF_KEY:
2438                         sref = (struct btrfs_shared_data_ref *)(iref + 1);
2439                         add_data_backref(extent_cache, key.objectid, offset,
2440                                         0, 0, 0,
2441                                         btrfs_shared_data_ref_count(eb, sref),
2442                                         0, key.offset);
2443                         break;
2444                 default:
2445                         fprintf(stderr, "corrupt extent record: key %Lu %u %Lu\n",
2446                                 key.objectid, key.type, key.offset);
2447                         goto out;
2448                 }
2449                 ptr += btrfs_extent_inline_ref_size(type);
2450         }
2451         WARN_ON(ptr > end);
2452 out:
2453         return 0;
2454 }
2455
2456 static int run_next_block(struct btrfs_root *root,
2457                           struct block_info *bits,
2458                           int bits_nr,
2459                           u64 *last,
2460                           struct cache_tree *pending,
2461                           struct cache_tree *seen,
2462                           struct cache_tree *reada,
2463                           struct cache_tree *nodes,
2464                           struct cache_tree *extent_cache)
2465 {
2466         struct extent_buffer *buf;
2467         u64 bytenr;
2468         u32 size;
2469         u64 parent;
2470         u64 owner;
2471         u64 flags;
2472         int ret;
2473         int i;
2474         int nritems;
2475         struct btrfs_key key;
2476         struct cache_extent *cache;
2477         int reada_bits;
2478
2479         ret = pick_next_pending(pending, reada, nodes, *last, bits,
2480                                 bits_nr, &reada_bits);
2481         if (ret == 0) {
2482                 return 1;
2483         }
2484         if (!reada_bits) {
2485                 for(i = 0; i < ret; i++) {
2486                         insert_cache_extent(reada, bits[i].start,
2487                                             bits[i].size);
2488
2489                         /* fixme, get the parent transid */
2490                         readahead_tree_block(root, bits[i].start,
2491                                              bits[i].size, 0);
2492                 }
2493         }
2494         *last = bits[0].start;
2495         bytenr = bits[0].start;
2496         size = bits[0].size;
2497
2498         cache = find_cache_extent(pending, bytenr, size);
2499         if (cache) {
2500                 remove_cache_extent(pending, cache);
2501                 free(cache);
2502         }
2503         cache = find_cache_extent(reada, bytenr, size);
2504         if (cache) {
2505                 remove_cache_extent(reada, cache);
2506                 free(cache);
2507         }
2508         cache = find_cache_extent(nodes, bytenr, size);
2509         if (cache) {
2510                 remove_cache_extent(nodes, cache);
2511                 free(cache);
2512         }
2513
2514         /* fixme, get the real parent transid */
2515         buf = read_tree_block(root, bytenr, size, 0);
2516         if (!extent_buffer_uptodate(buf)) {
2517                 record_bad_block_io(root->fs_info,
2518                                     extent_cache, bytenr, size);
2519                 free_extent_buffer(buf);
2520                 goto out;
2521         }
2522
2523         nritems = btrfs_header_nritems(buf);
2524
2525         ret = btrfs_lookup_extent_info(NULL, root, bytenr, size, NULL, &flags);
2526         if (ret < 0)
2527                 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
2528
2529         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
2530                 parent = bytenr;
2531                 owner = 0;
2532         } else {
2533                 parent = 0;
2534                 owner = btrfs_header_owner(buf);
2535         }
2536
2537         ret = check_block(root, extent_cache, buf, flags);
2538         if (ret)
2539                 goto out;
2540
2541         if (btrfs_is_leaf(buf)) {
2542                 btree_space_waste += btrfs_leaf_free_space(root, buf);
2543                 for (i = 0; i < nritems; i++) {
2544                         struct btrfs_file_extent_item *fi;
2545                         btrfs_item_key_to_cpu(buf, &key, i);
2546                         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
2547                                 process_extent_item(extent_cache, buf, i);
2548                                 continue;
2549                         }
2550                         if (key.type == BTRFS_EXTENT_CSUM_KEY) {
2551                                 total_csum_bytes +=
2552                                         btrfs_item_size_nr(buf, i);
2553                                 continue;
2554                         }
2555                         if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2556                                 continue;
2557                         }
2558                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
2559 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2560                                 process_extent_ref_v0(extent_cache, buf, i);
2561 #else
2562                                 BUG();
2563 #endif
2564                                 continue;
2565                         }
2566
2567                         if (key.type == BTRFS_TREE_BLOCK_REF_KEY) {
2568                                 add_tree_backref(extent_cache, key.objectid, 0,
2569                                                  key.offset, 0);
2570                                 continue;
2571                         }
2572                         if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2573                                 add_tree_backref(extent_cache, key.objectid,
2574                                                  key.offset, 0, 0);
2575                                 continue;
2576                         }
2577                         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
2578                                 struct btrfs_extent_data_ref *ref;
2579                                 ref = btrfs_item_ptr(buf, i,
2580                                                 struct btrfs_extent_data_ref);
2581                                 add_data_backref(extent_cache,
2582                                         key.objectid, 0,
2583                                         btrfs_extent_data_ref_root(buf, ref),
2584                                         btrfs_extent_data_ref_objectid(buf,
2585                                                                        ref),
2586                                         btrfs_extent_data_ref_offset(buf, ref),
2587                                         btrfs_extent_data_ref_count(buf, ref),
2588                                         0, root->sectorsize);
2589                                 continue;
2590                         }
2591                         if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
2592                                 struct btrfs_shared_data_ref *ref;
2593                                 ref = btrfs_item_ptr(buf, i,
2594                                                 struct btrfs_shared_data_ref);
2595                                 add_data_backref(extent_cache,
2596                                         key.objectid, key.offset, 0, 0, 0,
2597                                         btrfs_shared_data_ref_count(buf, ref),
2598                                         0, root->sectorsize);
2599                                 continue;
2600                         }
2601                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2602                                 continue;
2603                         fi = btrfs_item_ptr(buf, i,
2604                                             struct btrfs_file_extent_item);
2605                         if (btrfs_file_extent_type(buf, fi) ==
2606                             BTRFS_FILE_EXTENT_INLINE)
2607                                 continue;
2608                         if (btrfs_file_extent_disk_bytenr(buf, fi) == 0)
2609                                 continue;
2610
2611                         data_bytes_allocated +=
2612                                 btrfs_file_extent_disk_num_bytes(buf, fi);
2613                         if (data_bytes_allocated < root->sectorsize) {
2614                                 abort();
2615                         }
2616                         data_bytes_referenced +=
2617                                 btrfs_file_extent_num_bytes(buf, fi);
2618                         ret = add_extent_rec(extent_cache, NULL,
2619                                    btrfs_file_extent_disk_bytenr(buf, fi),
2620                                    btrfs_file_extent_disk_num_bytes(buf, fi),
2621                                    0, 0, 1, 1,
2622                                    btrfs_file_extent_disk_num_bytes(buf, fi));
2623                         add_data_backref(extent_cache,
2624                                 btrfs_file_extent_disk_bytenr(buf, fi),
2625                                 parent, owner, key.objectid, key.offset -
2626                                 btrfs_file_extent_offset(buf, fi), 1, 1,
2627                                 btrfs_file_extent_disk_num_bytes(buf, fi));
2628                         BUG_ON(ret);
2629                 }
2630         } else {
2631                 int level;
2632                 struct btrfs_key first_key;
2633
2634                 first_key.objectid = 0;
2635
2636                 if (nritems > 0)
2637                         btrfs_item_key_to_cpu(buf, &first_key, 0);
2638                 level = btrfs_header_level(buf);
2639                 for (i = 0; i < nritems; i++) {
2640                         u64 ptr = btrfs_node_blockptr(buf, i);
2641                         u32 size = btrfs_level_size(root, level - 1);
2642                         btrfs_node_key_to_cpu(buf, &key, i);
2643                         ret = add_extent_rec(extent_cache, &key,
2644                                              ptr, size, 0, 0, 1, 0, size);
2645                         BUG_ON(ret);
2646
2647                         add_tree_backref(extent_cache, ptr, parent, owner, 1);
2648
2649                         if (level > 1) {
2650                                 add_pending(nodes, seen, ptr, size);
2651                         } else {
2652                                 add_pending(pending, seen, ptr, size);
2653                         }
2654                 }
2655                 btree_space_waste += (BTRFS_NODEPTRS_PER_BLOCK(root) -
2656                                       nritems) * sizeof(struct btrfs_key_ptr);
2657         }
2658         total_btree_bytes += buf->len;
2659         if (fs_root_objectid(btrfs_header_owner(buf)))
2660                 total_fs_tree_bytes += buf->len;
2661         if (!found_old_backref &&
2662             btrfs_header_owner(buf) == BTRFS_TREE_RELOC_OBJECTID &&
2663             btrfs_header_backref_rev(buf) == BTRFS_MIXED_BACKREF_REV &&
2664             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC))
2665                 found_old_backref = 1;
2666 out:
2667         free_extent_buffer(buf);
2668         return 0;
2669 }
2670
2671 static int add_root_to_pending(struct extent_buffer *buf,
2672                                struct block_info *bits,
2673                                int bits_nr,
2674                                struct cache_tree *extent_cache,
2675                                struct cache_tree *pending,
2676                                struct cache_tree *seen,
2677                                struct cache_tree *reada,
2678                                struct cache_tree *nodes,
2679                                struct btrfs_key *root_key)
2680 {
2681         if (btrfs_header_level(buf) > 0)
2682                 add_pending(nodes, seen, buf->start, buf->len);
2683         else
2684                 add_pending(pending, seen, buf->start, buf->len);
2685         add_extent_rec(extent_cache, NULL, buf->start, buf->len,
2686                        0, 1, 1, 0, buf->len);
2687
2688         if (root_key->objectid == BTRFS_TREE_RELOC_OBJECTID ||
2689             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
2690                 add_tree_backref(extent_cache, buf->start, buf->start,
2691                                  0, 1);
2692         else
2693                 add_tree_backref(extent_cache, buf->start, 0,
2694                                  root_key->objectid, 1);
2695         return 0;
2696 }
2697
2698 /* as we fix the tree, we might be deleting blocks that
2699  * we're tracking for repair.  This hook makes sure we
2700  * remove any backrefs for blocks as we are fixing them.
2701  */
2702 static int free_extent_hook(struct btrfs_trans_handle *trans,
2703                             struct btrfs_root *root,
2704                             u64 bytenr, u64 num_bytes, u64 parent,
2705                             u64 root_objectid, u64 owner, u64 offset,
2706                             int refs_to_drop)
2707 {
2708         struct extent_record *rec;
2709         struct cache_extent *cache;
2710         int is_data;
2711         struct cache_tree *extent_cache = root->fs_info->fsck_extent_cache;
2712
2713         is_data = owner >= BTRFS_FIRST_FREE_OBJECTID;
2714         cache = find_cache_extent(extent_cache, bytenr, num_bytes);
2715         if (!cache)
2716                 return 0;
2717
2718         rec = container_of(cache, struct extent_record, cache);
2719         if (is_data) {
2720                 struct data_backref *back;
2721                 back = find_data_backref(rec, parent, root_objectid, owner,
2722                                          offset);
2723                 if (!back)
2724                         goto out;
2725                 if (back->node.found_ref) {
2726                         back->found_ref -= refs_to_drop;
2727                         if (rec->refs)
2728                                 rec->refs -= refs_to_drop;
2729                 }
2730                 if (back->node.found_extent_tree) {
2731                         back->num_refs -= refs_to_drop;
2732                         if (rec->extent_item_refs)
2733                                 rec->extent_item_refs -= refs_to_drop;
2734                 }
2735                 if (back->found_ref == 0)
2736                         back->node.found_ref = 0;
2737                 if (back->num_refs == 0)
2738                         back->node.found_extent_tree = 0;
2739
2740                 if (!back->node.found_extent_tree && back->node.found_ref) {
2741                         list_del(&back->node.list);
2742                         free(back);
2743                 }
2744         } else {
2745                 struct tree_backref *back;
2746                 back = find_tree_backref(rec, parent, root_objectid);
2747                 if (!back)
2748                         goto out;
2749                 if (back->node.found_ref) {
2750                         if (rec->refs)
2751                                 rec->refs--;
2752                         back->node.found_ref = 0;
2753                 }
2754                 if (back->node.found_extent_tree) {
2755                         if (rec->extent_item_refs)
2756                                 rec->extent_item_refs--;
2757                         back->node.found_extent_tree = 0;
2758                 }
2759                 if (!back->node.found_extent_tree && back->node.found_ref) {
2760                         list_del(&back->node.list);
2761                         free(back);
2762                 }
2763         }
2764         maybe_free_extent_rec(extent_cache, rec);
2765 out:
2766         return 0;
2767 }
2768
2769 static int delete_extent_records(struct btrfs_trans_handle *trans,
2770                                  struct btrfs_root *root,
2771                                  struct btrfs_path *path,
2772                                  u64 bytenr, u64 new_len)
2773 {
2774         struct btrfs_key key;
2775         struct btrfs_key found_key;
2776         struct extent_buffer *leaf;
2777         int ret;
2778         int slot;
2779
2780
2781         key.objectid = bytenr;
2782         key.type = (u8)-1;
2783         key.offset = (u64)-1;
2784
2785         while(1) {
2786                 ret = btrfs_search_slot(trans, root->fs_info->extent_root,
2787                                         &key, path, 0, 1);
2788                 if (ret < 0)
2789                         break;
2790
2791                 if (ret > 0) {
2792                         ret = 0;
2793                         if (path->slots[0] == 0)
2794                                 break;
2795                         path->slots[0]--;
2796                 }
2797                 ret = 0;
2798
2799                 leaf = path->nodes[0];
2800                 slot = path->slots[0];
2801
2802                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2803                 if (found_key.objectid != bytenr)
2804                         break;
2805
2806                 if (found_key.type != BTRFS_EXTENT_ITEM_KEY &&
2807                     found_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2808                     found_key.type != BTRFS_EXTENT_DATA_REF_KEY &&
2809                     found_key.type != BTRFS_EXTENT_REF_V0_KEY &&
2810                     found_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2811                     found_key.type != BTRFS_SHARED_DATA_REF_KEY) {
2812                         btrfs_release_path(NULL, path);
2813                         if (found_key.type == 0) {
2814                                 if (found_key.offset == 0)
2815                                         break;
2816                                 key.offset = found_key.offset - 1;
2817                                 key.type = found_key.type;
2818                         }
2819                         key.type = found_key.type - 1;
2820                         key.offset = (u64)-1;
2821                         continue;
2822                 }
2823
2824                 fprintf(stderr, "repair deleting extent record: key %Lu %u %Lu\n",
2825                         found_key.objectid, found_key.type, found_key.offset);
2826
2827                 ret = btrfs_del_item(trans, root->fs_info->extent_root, path);
2828                 if (ret)
2829                         break;
2830                 btrfs_release_path(NULL, path);
2831
2832                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY) {
2833                         ret = btrfs_update_block_group(trans, root, bytenr,
2834                                                        found_key.offset, 0, 0);
2835                         if (ret)
2836                                 break;
2837                 }
2838         }
2839
2840         btrfs_release_path(NULL, path);
2841         return ret;
2842 }
2843
2844 /*
2845  * for a single backref, this will allocate a new extent
2846  * and add the backref to it.
2847  */
2848 static int record_extent(struct btrfs_trans_handle *trans,
2849                          struct btrfs_fs_info *info,
2850                          struct btrfs_path *path,
2851                          struct extent_record *rec,
2852                          struct extent_backref *back,
2853                          int allocated, u64 flags)
2854 {
2855         int ret;
2856         struct btrfs_root *extent_root = info->extent_root;
2857         struct extent_buffer *leaf;
2858         struct btrfs_key ins_key;
2859         struct btrfs_extent_item *ei;
2860         struct tree_backref *tback;
2861         struct data_backref *dback;
2862         struct btrfs_tree_block_info *bi;
2863
2864         if (!back->is_data)
2865                 rec->max_size = max_t(u64, rec->max_size,
2866                                     info->extent_root->leafsize);
2867
2868         if (!allocated) {
2869                 u32 item_size = sizeof(*ei);
2870
2871                 if (!back->is_data)
2872                         item_size += sizeof(*bi);
2873
2874                 ins_key.objectid = rec->start;
2875                 ins_key.offset = rec->max_size;
2876                 ins_key.type = BTRFS_EXTENT_ITEM_KEY;
2877
2878                 ret = btrfs_insert_empty_item(trans, extent_root, path,
2879                                         &ins_key, item_size);
2880                 if (ret)
2881                         goto fail;
2882
2883                 leaf = path->nodes[0];
2884                 ei = btrfs_item_ptr(leaf, path->slots[0],
2885                                     struct btrfs_extent_item);
2886
2887                 btrfs_set_extent_refs(leaf, ei, 0);
2888                 btrfs_set_extent_generation(leaf, ei, rec->generation);
2889
2890                 if (back->is_data) {
2891                         btrfs_set_extent_flags(leaf, ei,
2892                                                BTRFS_EXTENT_FLAG_DATA);
2893                 } else {
2894                         struct btrfs_disk_key copy_key;;
2895
2896                         tback = (struct tree_backref *)back;
2897                         bi = (struct btrfs_tree_block_info *)(ei + 1);
2898                         memset_extent_buffer(leaf, 0, (unsigned long)bi,
2899                                              sizeof(*bi));
2900                         memset(&copy_key, 0, sizeof(copy_key));
2901
2902                         copy_key.objectid = le64_to_cpu(rec->info_objectid);
2903                         btrfs_set_tree_block_level(leaf, bi, rec->info_level);
2904                         btrfs_set_tree_block_key(leaf, bi, &copy_key);
2905
2906                         btrfs_set_extent_flags(leaf, ei,
2907                                                BTRFS_EXTENT_FLAG_TREE_BLOCK | flags);
2908                 }
2909
2910                 btrfs_mark_buffer_dirty(leaf);
2911                 ret = btrfs_update_block_group(trans, extent_root, rec->start,
2912                                                rec->max_size, 1, 0);
2913                 if (ret)
2914                         goto fail;
2915                 btrfs_release_path(NULL, path);
2916         }
2917
2918         if (back->is_data) {
2919                 u64 parent;
2920                 int i;
2921
2922                 dback = (struct data_backref *)back;
2923                 if (back->full_backref)
2924                         parent = dback->parent;
2925                 else
2926                         parent = 0;
2927
2928                 for (i = 0; i < dback->found_ref; i++) {
2929                         /* if parent != 0, we're doing a full backref
2930                          * passing BTRFS_FIRST_FREE_OBJECTID as the owner
2931                          * just makes the backref allocator create a data
2932                          * backref
2933                          */
2934                         ret = btrfs_inc_extent_ref(trans, info->extent_root,
2935                                                    rec->start, rec->max_size,
2936                                                    parent,
2937                                                    dback->root,
2938                                                    parent ?
2939                                                    BTRFS_FIRST_FREE_OBJECTID :
2940                                                    dback->owner,
2941                                                    dback->offset);
2942                         if (ret)
2943                                 break;
2944                 }
2945                 fprintf(stderr, "adding new data backref"
2946                                 " on %llu %s %llu owner %llu"
2947                                 " offset %llu found %d\n",
2948                                 (unsigned long long)rec->start,
2949                                 back->full_backref ?
2950                                 "parent" : "root",
2951                                 back->full_backref ?
2952                                 (unsigned long long)parent :
2953                                 (unsigned long long)dback->root,
2954                                 (unsigned long long)dback->owner,
2955                                 (unsigned long long)dback->offset,
2956                                 dback->found_ref);
2957         } else {
2958                 u64 parent;
2959
2960                 tback = (struct tree_backref *)back;
2961                 if (back->full_backref)
2962                         parent = tback->parent;
2963                 else
2964                         parent = 0;
2965
2966                 ret = btrfs_inc_extent_ref(trans, info->extent_root,
2967                                            rec->start, rec->max_size,
2968                                            parent, tback->root, 0, 0);
2969                 fprintf(stderr, "adding new tree backref on "
2970                         "start %llu len %llu parent %llu root %llu\n",
2971                         rec->start, rec->max_size, tback->parent, tback->root);
2972         }
2973         if (ret)
2974                 goto fail;
2975 fail:
2976         btrfs_release_path(NULL, path);
2977         return ret;
2978 }
2979
2980 /*
2981  * when an incorrect extent item is found, this will delete
2982  * all of the existing entries for it and recreate them
2983  * based on what the tree scan found.
2984  */
2985 static int fixup_extent_refs(struct btrfs_trans_handle *trans,
2986                              struct btrfs_fs_info *info,
2987                              struct extent_record *rec)
2988 {
2989         int ret;
2990         struct btrfs_path *path;
2991         struct list_head *cur = rec->backrefs.next;
2992         struct cache_extent *cache;
2993         struct extent_backref *back;
2994         int allocated = 0;
2995         u64 flags = 0;
2996
2997         /* remember our flags for recreating the extent */
2998         ret = btrfs_lookup_extent_info(NULL, info->extent_root, rec->start,
2999                                        rec->max_size, NULL, &flags);
3000         if (ret < 0)
3001                 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
3002
3003         path = btrfs_alloc_path();
3004
3005         /* step one, delete all the existing records */
3006         ret = delete_extent_records(trans, info->extent_root, path,
3007                                     rec->start, rec->max_size);
3008
3009         if (ret < 0)
3010                 goto out;
3011
3012         /* was this block corrupt?  If so, don't add references to it */
3013         cache = find_cache_extent(info->corrupt_blocks, rec->start, rec->max_size);
3014         if (cache) {
3015                 ret = 0;
3016                 goto out;
3017         }
3018
3019         /* step two, recreate all the refs we did find */
3020         while(cur != &rec->backrefs) {
3021                 back = list_entry(cur, struct extent_backref, list);
3022                 cur = cur->next;
3023
3024                 /*
3025                  * if we didn't find any references, don't create a
3026                  * new extent record
3027                  */
3028                 if (!back->found_ref)
3029                         continue;
3030
3031                 ret = record_extent(trans, info, path, rec, back, allocated, flags);
3032                 allocated = 1;
3033
3034                 if (ret)
3035                         goto out;
3036         }
3037 out:
3038         btrfs_free_path(path);
3039         return ret;
3040 }
3041
3042 /* right now we only prune from the extent allocation tree */
3043 static int prune_one_block(struct btrfs_trans_handle *trans,
3044                            struct btrfs_fs_info *info,
3045                            struct btrfs_corrupt_block *corrupt)
3046 {
3047         int ret;
3048         struct btrfs_path path;
3049         struct extent_buffer *eb;
3050         u64 found;
3051         int slot;
3052         int nritems;
3053         int level = corrupt->level + 1;
3054
3055         btrfs_init_path(&path);
3056 again:
3057         /* we want to stop at the parent to our busted block */
3058         path.lowest_level = level;
3059
3060         ret = btrfs_search_slot(trans, info->extent_root,
3061                                 &corrupt->key, &path, -1, 1);
3062
3063         if (ret < 0)
3064                 goto out;
3065
3066         eb = path.nodes[level];
3067         if (!eb) {
3068                 ret = -ENOENT;
3069                 goto out;
3070         }
3071
3072         /*
3073          * hopefully the search gave us the block we want to prune,
3074          * lets try that first
3075          */
3076         slot = path.slots[level];
3077         found =  btrfs_node_blockptr(eb, slot);
3078         if (found == corrupt->cache.start)
3079                 goto del_ptr;
3080
3081         nritems = btrfs_header_nritems(eb);
3082
3083         /* the search failed, lets scan this node and hope we find it */
3084         for (slot = 0; slot < nritems; slot++) {
3085                 found =  btrfs_node_blockptr(eb, slot);
3086                 if (found == corrupt->cache.start)
3087                         goto del_ptr;
3088         }
3089         /*
3090          * we couldn't find the bad block.  TODO, search all the nodes for pointers
3091          * to this block
3092          */
3093         if (eb == info->extent_root->node) {
3094                 ret = -ENOENT;
3095                 goto out;
3096         } else {
3097                 level++;
3098                 btrfs_release_path(NULL, &path);
3099                 goto again;
3100         }
3101
3102 del_ptr:
3103         printk("deleting pointer to block %Lu\n", corrupt->cache.start);
3104         ret = btrfs_del_ptr(trans, info->extent_root, &path, level, slot);
3105
3106 out:
3107         btrfs_release_path(NULL, &path);
3108         return ret;
3109 }
3110
3111 static int prune_corrupt_blocks(struct btrfs_trans_handle *trans,
3112                                 struct btrfs_fs_info *info)
3113 {
3114         struct cache_extent *cache;
3115         struct btrfs_corrupt_block *corrupt;
3116
3117         cache = find_first_cache_extent(info->corrupt_blocks, 0);
3118         while (1) {
3119                 if (!cache)
3120                         break;
3121                 corrupt = container_of(cache, struct btrfs_corrupt_block, cache);
3122                 prune_one_block(trans, info, corrupt);
3123                 cache = next_cache_extent(cache);
3124         }
3125         return 0;
3126 }
3127
3128 static void free_corrupt_blocks(struct btrfs_fs_info *info)
3129 {
3130         struct cache_extent *cache;
3131         struct btrfs_corrupt_block *corrupt;
3132
3133         while (1) {
3134                 cache = find_first_cache_extent(info->corrupt_blocks, 0);
3135                 if (!cache)
3136                         break;
3137                 corrupt = container_of(cache, struct btrfs_corrupt_block, cache);
3138                 remove_cache_extent(info->corrupt_blocks, cache);
3139                 free(corrupt);
3140         }
3141 }
3142
3143 static int check_extent_refs(struct btrfs_trans_handle *trans,
3144                              struct btrfs_root *root,
3145                              struct cache_tree *extent_cache, int repair)
3146 {
3147         struct extent_record *rec;
3148         struct cache_extent *cache;
3149         int err = 0;
3150         int ret = 0;
3151         int fixed = 0;
3152
3153         if (repair) {
3154                 /*
3155                  * if we're doing a repair, we have to make sure
3156                  * we don't allocate from the problem extents.
3157                  * In the worst case, this will be all the
3158                  * extents in the FS
3159                  */
3160                 cache = find_first_cache_extent(extent_cache, 0);
3161                 while(cache) {
3162                         rec = container_of(cache, struct extent_record, cache);
3163                         btrfs_pin_extent(root->fs_info,
3164                                          rec->start, rec->max_size);
3165                         cache = next_cache_extent(cache);
3166                 }
3167
3168                 /* pin down all the corrupted blocks too */
3169                 cache = find_first_cache_extent(root->fs_info->corrupt_blocks, 0);
3170                 while(cache) {
3171                         rec = container_of(cache, struct extent_record, cache);
3172                         btrfs_pin_extent(root->fs_info,
3173                                          rec->start, rec->max_size);
3174                         cache = next_cache_extent(cache);
3175                 }
3176                 prune_corrupt_blocks(trans, root->fs_info);
3177         }
3178         while(1) {
3179                 fixed = 0;
3180                 cache = find_first_cache_extent(extent_cache, 0);
3181                 if (!cache)
3182                         break;
3183                 rec = container_of(cache, struct extent_record, cache);
3184                 if (rec->refs != rec->extent_item_refs) {
3185                         fprintf(stderr, "ref mismatch on [%llu %llu] ",
3186                                 (unsigned long long)rec->start,
3187                                 (unsigned long long)rec->nr);
3188                         fprintf(stderr, "extent item %llu, found %llu\n",
3189                                 (unsigned long long)rec->extent_item_refs,
3190                                 (unsigned long long)rec->refs);
3191                         if (!fixed && repair) {
3192                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3193                                 if (ret)
3194                                         goto repair_abort;
3195                                 fixed = 1;
3196                         }
3197                         err = 1;
3198
3199                 }
3200                 if (all_backpointers_checked(rec, 1)) {
3201                         fprintf(stderr, "backpointer mismatch on [%llu %llu]\n",
3202                                 (unsigned long long)rec->start,
3203                                 (unsigned long long)rec->nr);
3204
3205                         if (!fixed && repair) {
3206                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3207                                 if (ret)
3208                                         goto repair_abort;
3209                                 fixed = 1;
3210                         }
3211
3212                         err = 1;
3213                 }
3214                 if (!rec->owner_ref_checked) {
3215                         fprintf(stderr, "owner ref check failed [%llu %llu]\n",
3216                                 (unsigned long long)rec->start,
3217                                 (unsigned long long)rec->nr);
3218                         if (!fixed && repair) {
3219                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3220                                 if (ret)
3221                                         goto repair_abort;
3222                                 fixed = 1;
3223                         }
3224                         err = 1;
3225                 }
3226
3227                 remove_cache_extent(extent_cache, cache);
3228                 free_all_extent_backrefs(rec);
3229                 free(rec);
3230         }
3231 repair_abort:
3232         if (repair) {
3233                 if (ret) {
3234                         fprintf(stderr, "failed to repair damaged filesystem, aborting\n");
3235                         exit(1);
3236                 } else {
3237                         btrfs_fix_block_accounting(trans, root);
3238                 }
3239                 if (err)
3240                         fprintf(stderr, "repaired damaged extent references\n");
3241                 return ret;
3242         }
3243         return err;
3244 }
3245
3246 static int check_extents(struct btrfs_trans_handle *trans,
3247                          struct btrfs_root *root, int repair)
3248 {
3249         struct cache_tree extent_cache;
3250         struct cache_tree seen;
3251         struct cache_tree pending;
3252         struct cache_tree reada;
3253         struct cache_tree nodes;
3254         struct cache_tree corrupt_blocks;
3255         struct btrfs_path path;
3256         struct btrfs_key key;
3257         struct btrfs_key found_key;
3258         int ret;
3259         u64 last = 0;
3260         struct block_info *bits;
3261         int bits_nr;
3262         struct extent_buffer *leaf;
3263         int slot;
3264         struct btrfs_root_item ri;
3265
3266         cache_tree_init(&extent_cache);
3267         cache_tree_init(&seen);
3268         cache_tree_init(&pending);
3269         cache_tree_init(&nodes);
3270         cache_tree_init(&reada);
3271         cache_tree_init(&corrupt_blocks);
3272
3273         if (repair) {
3274                 root->fs_info->fsck_extent_cache = &extent_cache;
3275                 root->fs_info->free_extent_hook = free_extent_hook;
3276                 root->fs_info->corrupt_blocks = &corrupt_blocks;
3277         }
3278
3279         bits_nr = 1024;
3280         bits = malloc(bits_nr * sizeof(struct block_info));
3281         if (!bits) {
3282                 perror("malloc");
3283                 exit(1);
3284         }
3285
3286         add_root_to_pending(root->fs_info->tree_root->node, bits, bits_nr,
3287                             &extent_cache, &pending, &seen, &reada, &nodes,
3288                             &root->fs_info->tree_root->root_key);
3289
3290         add_root_to_pending(root->fs_info->chunk_root->node, bits, bits_nr,
3291                             &extent_cache, &pending, &seen, &reada, &nodes,
3292                             &root->fs_info->chunk_root->root_key);
3293
3294         btrfs_init_path(&path);
3295         key.offset = 0;
3296         key.objectid = 0;
3297         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
3298         ret = btrfs_search_slot(NULL, root->fs_info->tree_root,
3299                                         &key, &path, 0, 0);
3300         BUG_ON(ret < 0);
3301         while(1) {
3302                 leaf = path.nodes[0];
3303                 slot = path.slots[0];
3304                 if (slot >= btrfs_header_nritems(path.nodes[0])) {
3305                         ret = btrfs_next_leaf(root, &path);
3306                         if (ret != 0)
3307                                 break;
3308                         leaf = path.nodes[0];
3309                         slot = path.slots[0];
3310                 }
3311                 btrfs_item_key_to_cpu(leaf, &found_key, path.slots[0]);
3312                 if (btrfs_key_type(&found_key) == BTRFS_ROOT_ITEM_KEY) {
3313                         unsigned long offset;
3314                         struct extent_buffer *buf;
3315
3316                         offset = btrfs_item_ptr_offset(leaf, path.slots[0]);
3317                         read_extent_buffer(leaf, &ri, offset, sizeof(ri));
3318                         buf = read_tree_block(root->fs_info->tree_root,
3319                                               btrfs_root_bytenr(&ri),
3320                                               btrfs_level_size(root,
3321                                                btrfs_root_level(&ri)), 0);
3322                         add_root_to_pending(buf, bits, bits_nr, &extent_cache,
3323                                             &pending, &seen, &reada, &nodes,
3324                                             &found_key);
3325                         free_extent_buffer(buf);
3326                 }
3327                 path.slots[0]++;
3328         }
3329         btrfs_release_path(root, &path);
3330         while(1) {
3331                 ret = run_next_block(root, bits, bits_nr, &last, &pending,
3332                                      &seen, &reada, &nodes, &extent_cache);
3333                 if (ret != 0)
3334                         break;
3335         }
3336         ret = check_extent_refs(trans, root, &extent_cache, repair);
3337
3338         if (repair) {
3339                 free_corrupt_blocks(root->fs_info);
3340                 root->fs_info->fsck_extent_cache = NULL;
3341                 root->fs_info->free_extent_hook = NULL;
3342                 root->fs_info->corrupt_blocks = NULL;
3343         }
3344
3345         return ret;
3346 }
3347
3348 static void print_usage(void)
3349 {
3350         fprintf(stderr, "usage: btrfsck dev\n");
3351         fprintf(stderr, "%s\n", BTRFS_BUILD_VERSION);
3352         exit(1);
3353 }
3354
3355 static struct option long_options[] = {
3356         { "super", 1, NULL, 's' },
3357         { "repair", 0, NULL, 0 },
3358         { "init-csum-tree", 0, NULL, 0 },
3359         { 0, 0, 0, 0}
3360 };
3361
3362 int main(int ac, char **av)
3363 {
3364         struct cache_tree root_cache;
3365         struct btrfs_root *root;
3366         struct btrfs_fs_info *info;
3367         struct btrfs_trans_handle *trans = NULL;
3368         u64 bytenr = 0;
3369         int ret;
3370         int num;
3371         int repair = 0;
3372         int option_index = 0;
3373         int init_csum_tree = 0;
3374         int rw = 0;
3375
3376         while(1) {
3377                 int c;
3378                 c = getopt_long(ac, av, "", long_options,
3379                                 &option_index);
3380                 if (c < 0)
3381                         break;
3382                 switch(c) {
3383                         case 's':
3384                                 num = atol(optarg);
3385                                 bytenr = btrfs_sb_offset(num);
3386                                 printf("using SB copy %d, bytenr %llu\n", num,
3387                                        (unsigned long long)bytenr);
3388                                 break;
3389                         case '?':
3390                                 print_usage();
3391                 }
3392                 if (option_index == 1) {
3393                         printf("enabling repair mode\n");
3394                         repair = 1;
3395                         rw = 1;
3396                 } else if (option_index == 2) {
3397                         printf("Creating a new CRC tree\n");
3398                         init_csum_tree = 1;
3399                         rw = 1;
3400                 }
3401
3402         }
3403         ac = ac - optind;
3404
3405         if (ac != 1)
3406                 print_usage();
3407
3408         radix_tree_init();
3409         cache_tree_init(&root_cache);
3410
3411         if((ret = check_mounted(av[optind])) < 0) {
3412                 fprintf(stderr, "Could not check mount status: %s\n", strerror(-ret));
3413                 return ret;
3414         } else if(ret) {
3415                 fprintf(stderr, "%s is currently mounted. Aborting.\n", av[optind]);
3416                 return -EBUSY;
3417         }
3418
3419         info = open_ctree_fs_info(av[optind], bytenr, rw, 1);
3420
3421         if (info == NULL)
3422                 return 1;
3423
3424         if (!extent_buffer_uptodate(info->tree_root->node) ||
3425             !extent_buffer_uptodate(info->dev_root->node) ||
3426             !extent_buffer_uptodate(info->extent_root->node) ||
3427             !extent_buffer_uptodate(info->chunk_root->node)) {
3428                 fprintf(stderr, "Critical roots corrupted, unable to fsck the FS\n");
3429                 return -EIO;
3430         }
3431
3432         root = info->fs_root;
3433
3434         fprintf(stderr, "checking extents\n");
3435         if (rw)
3436                 trans = btrfs_start_transaction(root, 1);
3437
3438         if (init_csum_tree) {
3439                 fprintf(stderr, "Reinit crc root\n");
3440                 ret = btrfs_fsck_reinit_root(trans, info->csum_root);
3441                 if (ret) {
3442                         fprintf(stderr, "crc root initialization failed\n");
3443                         return -EIO;
3444                 }
3445                 goto out;
3446         }
3447
3448         ret = check_extents(trans, root, repair);
3449         if (ret)
3450                 fprintf(stderr, "Errors found in extent allocation tree\n");
3451
3452         fprintf(stderr, "checking fs roots\n");
3453         ret = check_fs_roots(root, &root_cache);
3454         if (ret)
3455                 goto out;
3456
3457         fprintf(stderr, "checking root refs\n");
3458         ret = check_root_refs(root, &root_cache);
3459 out:
3460         free_root_recs(&root_cache);
3461         if (rw) {
3462                 ret = btrfs_commit_transaction(trans, root);
3463                 if (ret)
3464                         exit(1);
3465         }
3466         close_ctree(root);
3467
3468         if (found_old_backref) { /*
3469                  * there was a disk format change when mixed
3470                  * backref was in testing tree. The old format
3471                  * existed about one week.
3472                  */
3473                 printf("\n * Found old mixed backref format. "
3474                        "The old format is not supported! *"
3475                        "\n * Please mount the FS in readonly mode, "
3476                        "backup data and re-format the FS. *\n\n");
3477                 ret = 1;
3478         }
3479         printf("found %llu bytes used err is %d\n",
3480                (unsigned long long)bytes_used, ret);
3481         printf("total csum bytes: %llu\n",(unsigned long long)total_csum_bytes);
3482         printf("total tree bytes: %llu\n",
3483                (unsigned long long)total_btree_bytes);
3484         printf("total fs tree bytes: %llu\n",
3485                (unsigned long long)total_fs_tree_bytes);
3486         printf("btree space waste bytes: %llu\n",
3487                (unsigned long long)btree_space_waste);
3488         printf("file data blocks allocated: %llu\n referenced %llu\n",
3489                 (unsigned long long)data_bytes_allocated,
3490                 (unsigned long long)data_bytes_referenced);
3491         printf("%s\n", BTRFS_BUILD_VERSION);
3492         return ret;
3493 }