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 "volumes.h"
  30 #include "repair.h"
  31 #include "disk-io.h"
  32 #include "print-tree.h"
  33 #include "transaction.h"
  34 #include "list.h"
  35 #include "version.h"
  36 #include "utils.h"
  37
  38 static u64 bytes_used = 0;
  39 static u64 total_csum_bytes = 0;
  40 static u64 total_btree_bytes = 0;
  41 static u64 total_fs_tree_bytes = 0;
  42 static u64 btree_space_waste = 0;
  43 static u64 data_bytes_allocated = 0;
  44 static u64 data_bytes_referenced = 0;
  45 static int found_old_backref = 0;
  46
  47 struct extent_backref {
  48         struct list_head list;
  49         unsigned int is_data:1;
  50         unsigned int found_extent_tree:1;
  51         unsigned int full_backref:1;
  52         unsigned int found_ref:1;
  53 };
  54
  55 struct data_backref {
  56         struct extent_backref node;
  57         union {
  58                 u64 parent;
  59                 u64 root;
  60         };
  61         u64 owner;
  62         u64 offset;
  63         u32 num_refs;
  64         u32 found_ref;
  65 };
  66
  67 struct tree_backref {
  68         struct extent_backref node;
  69         union {
  70                 u64 parent;
  71                 u64 root;
  72         };
  73 };
  74
  75 struct extent_record {
  76         struct list_head backrefs;
  77         struct cache_extent cache;
  78         struct btrfs_disk_key parent_key;
  79         u64 start;
  80         u64 max_size;
  81         u64 nr;
  82         u64 refs;
  83         u64 extent_item_refs;
  84         u64 generation;
  85         u64 info_objectid;
  86         u8 info_level;
  87         unsigned int content_checked:1;
  88         unsigned int owner_ref_checked:1;
  89         unsigned int is_root:1;
  90 };
  91
  92 struct inode_backref {
  93         struct list_head list;
  94         unsigned int found_dir_item:1;
  95         unsigned int found_dir_index:1;
  96         unsigned int found_inode_ref:1;
  97         unsigned int filetype:8;
  98         int errors;
  99         u64 dir;
 100         u64 index;
 101         u16 namelen;
 102         char name[0];
 103 };
 104
 105 #define REF_ERR_NO_DIR_ITEM             (1 << 0)
 106 #define REF_ERR_NO_DIR_INDEX            (1 << 1)
 107 #define REF_ERR_NO_INODE_REF            (1 << 2)
 108 #define REF_ERR_DUP_DIR_ITEM            (1 << 3)
 109 #define REF_ERR_DUP_DIR_INDEX           (1 << 4)
 110 #define REF_ERR_DUP_INODE_REF           (1 << 5)
 111 #define REF_ERR_INDEX_UNMATCH           (1 << 6)
 112 #define REF_ERR_FILETYPE_UNMATCH        (1 << 7)
 113 #define REF_ERR_NAME_TOO_LONG           (1 << 8) // 100
 114 #define REF_ERR_NO_ROOT_REF             (1 << 9)
 115 #define REF_ERR_NO_ROOT_BACKREF         (1 << 10)
 116 #define REF_ERR_DUP_ROOT_REF            (1 << 11)
 117 #define REF_ERR_DUP_ROOT_BACKREF        (1 << 12)
 118
 119 struct inode_record {
 120         struct list_head backrefs;
 121         unsigned int checked:1;
 122         unsigned int merging:1;
 123         unsigned int found_inode_item:1;
 124         unsigned int found_dir_item:1;
 125         unsigned int found_file_extent:1;
 126         unsigned int found_csum_item:1;
 127         unsigned int some_csum_missing:1;
 128         unsigned int nodatasum:1;
 129         int errors;
 130
 131         u64 ino;
 132         u32 nlink;
 133         u32 imode;
 134         u64 isize;
 135         u64 nbytes;
 136
 137         u32 found_link;
 138         u64 found_size;
 139         u64 extent_start;
 140         u64 extent_end;
 141         u64 first_extent_gap;
 142
 143         u32 refs;
 144 };
 145
 146 #define I_ERR_NO_INODE_ITEM             (1 << 0)
 147 #define I_ERR_NO_ORPHAN_ITEM            (1 << 1)
 148 #define I_ERR_DUP_INODE_ITEM            (1 << 2)
 149 #define I_ERR_DUP_DIR_INDEX             (1 << 3)
 150 #define I_ERR_ODD_DIR_ITEM              (1 << 4)
 151 #define I_ERR_ODD_FILE_EXTENT           (1 << 5)
 152 #define I_ERR_BAD_FILE_EXTENT           (1 << 6)
 153 #define I_ERR_FILE_EXTENT_OVERLAP       (1 << 7)
 154 #define I_ERR_FILE_EXTENT_DISCOUNT      (1 << 8) // 100
 155 #define I_ERR_DIR_ISIZE_WRONG           (1 << 9)
 156 #define I_ERR_FILE_NBYTES_WRONG         (1 << 10) // 400
 157 #define I_ERR_ODD_CSUM_ITEM             (1 << 11)
 158 #define I_ERR_SOME_CSUM_MISSING         (1 << 12)
 159 #define I_ERR_LINK_COUNT_WRONG          (1 << 13)
 160
 161 struct root_backref {
 162         struct list_head list;
 163         unsigned int found_dir_item:1;
 164         unsigned int found_dir_index:1;
 165         unsigned int found_back_ref:1;
 166         unsigned int found_forward_ref:1;
 167         unsigned int reachable:1;
 168         int errors;
 169         u64 ref_root;
 170         u64 dir;
 171         u64 index;
 172         u16 namelen;
 173         char name[0];
 174 };
 175
 176 struct root_record {
 177         struct list_head backrefs;
 178         struct cache_extent cache;
 179         unsigned int found_root_item:1;
 180         u64 objectid;
 181         u32 found_ref;
 182 };
 183
 184 struct ptr_node {
 185         struct cache_extent cache;
 186         void *data;
 187 };
 188
 189 struct shared_node {
 190         struct cache_extent cache;
 191         struct cache_tree root_cache;
 192         struct cache_tree inode_cache;
 193         struct inode_record *current;
 194         u32 refs;
 195 };
 196
 197 struct block_info {
 198         u64 start;
 199         u32 size;
 200 };
 201
 202 struct walk_control {
 203         struct cache_tree shared;
 204         struct shared_node *nodes[BTRFS_MAX_LEVEL];
 205         int active_node;
 206         int root_level;
 207 };
 208
 209 static u8 imode_to_type(u32 imode)
 210 {
 211 #define S_SHIFT 12
 212         static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
 213                 [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
 214                 [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
 215                 [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
 216                 [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
 217                 [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 218                 [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 219                 [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 220         };
 221
 222         return btrfs_type_by_mode[(imode & S_IFMT) >> S_SHIFT];
 223 #undef S_SHIFT
 224 }
 225
 226 static struct inode_record *clone_inode_rec(struct inode_record *orig_rec)
 227 {
 228         struct inode_record *rec;
 229         struct inode_backref *backref;
 230         struct inode_backref *orig;
 231         size_t size;
 232
 233         rec = malloc(sizeof(*rec));
 234         memcpy(rec, orig_rec, sizeof(*rec));
 235         rec->refs = 1;
 236         INIT_LIST_HEAD(&rec->backrefs);
 237
 238         list_for_each_entry(orig, &orig_rec->backrefs, list) {
 239                 size = sizeof(*orig) + orig->namelen + 1;
 240                 backref = malloc(size);
 241                 memcpy(backref, orig, size);
 242                 list_add_tail(&backref->list, &rec->backrefs);
 243         }
 244         return rec;
 245 }
 246
 247 static struct inode_record *get_inode_rec(struct cache_tree *inode_cache,
 248                                           u64 ino, int mod)
 249 {
 250         struct ptr_node *node;
 251         struct cache_extent *cache;
 252         struct inode_record *rec = NULL;
 253         int ret;
 254
 255         cache = find_cache_extent(inode_cache, ino, 1);
 256         if (cache) {
 257                 node = container_of(cache, struct ptr_node, cache);
 258                 rec = node->data;
 259                 if (mod && rec->refs > 1) {
 260                         node->data = clone_inode_rec(rec);
 261                         rec->refs--;
 262                         rec = node->data;
 263                 }
 264         } else if (mod) {
 265                 rec = calloc(1, sizeof(*rec));
 266                 rec->ino = ino;
 267                 rec->extent_start = (u64)-1;
 268                 rec->first_extent_gap = (u64)-1;
 269                 rec->refs = 1;
 270                 INIT_LIST_HEAD(&rec->backrefs);
 271
 272                 node = malloc(sizeof(*node));
 273                 node->cache.start = ino;
 274                 node->cache.size = 1;
 275                 node->data = rec;
 276
 277                 if (ino == BTRFS_FREE_INO_OBJECTID)
 278                         rec->found_link = 1;
 279
 280                 ret = insert_existing_cache_extent(inode_cache, &node->cache);
 281                 BUG_ON(ret);
 282         }
 283         return rec;
 284 }
 285
 286 static void free_inode_rec(struct inode_record *rec)
 287 {
 288         struct inode_backref *backref;
 289
 290         if (--rec->refs > 0)
 291                 return;
 292
 293         while (!list_empty(&rec->backrefs)) {
 294                 backref = list_entry(rec->backrefs.next,
 295                                      struct inode_backref, list);
 296                 list_del(&backref->list);
 297                 free(backref);
 298         }
 299         free(rec);
 300 }
 301
 302 static int can_free_inode_rec(struct inode_record *rec)
 303 {
 304         if (!rec->errors && rec->checked && rec->found_inode_item &&
 305             rec->nlink == rec->found_link && list_empty(&rec->backrefs))
 306                 return 1;
 307         return 0;
 308 }
 309
 310 static void maybe_free_inode_rec(struct cache_tree *inode_cache,
 311                                  struct inode_record *rec)
 312 {
 313         struct cache_extent *cache;
 314         struct inode_backref *tmp, *backref;
 315         struct ptr_node *node;
 316         unsigned char filetype;
 317
 318         if (!rec->found_inode_item)
 319                 return;
 320
 321         filetype = imode_to_type(rec->imode);
 322         list_for_each_entry_safe(backref, tmp, &rec->backrefs, list) {
 323                 if (backref->found_dir_item && backref->found_dir_index) {
 324                         if (backref->filetype != filetype)
 325                                 backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 326                         if (!backref->errors && backref->found_inode_ref) {
 327                                 list_del(&backref->list);
 328                                 free(backref);
 329                         }
 330                 }
 331         }
 332
 333         if (!rec->checked || rec->merging)
 334                 return;
 335
 336         if (S_ISDIR(rec->imode)) {
 337                 if (rec->found_size != rec->isize)
 338                         rec->errors |= I_ERR_DIR_ISIZE_WRONG;
 339                 if (rec->found_file_extent)
 340                         rec->errors |= I_ERR_ODD_FILE_EXTENT;
 341         } else if (S_ISREG(rec->imode) || S_ISLNK(rec->imode)) {
 342                 if (rec->found_dir_item)
 343                         rec->errors |= I_ERR_ODD_DIR_ITEM;
 344                 if (rec->found_size != rec->nbytes)
 345                         rec->errors |= I_ERR_FILE_NBYTES_WRONG;
 346                 if (rec->extent_start == (u64)-1 || rec->extent_start > 0)
 347                         rec->first_extent_gap = 0;
 348                 if (rec->nlink > 0 && (rec->extent_end < rec->isize ||
 349                     rec->first_extent_gap < rec->isize))
 350                         rec->errors |= I_ERR_FILE_EXTENT_DISCOUNT;
 351         }
 352
 353         if (S_ISREG(rec->imode) || S_ISLNK(rec->imode)) {
 354                 if (rec->found_csum_item && rec->nodatasum)
 355                         rec->errors |= I_ERR_ODD_CSUM_ITEM;
 356                 if (rec->some_csum_missing && !rec->nodatasum)
 357                         rec->errors |= I_ERR_SOME_CSUM_MISSING;
 358         }
 359
 360         BUG_ON(rec->refs != 1);
 361         if (can_free_inode_rec(rec)) {
 362                 cache = find_cache_extent(inode_cache, rec->ino, 1);
 363                 node = container_of(cache, struct ptr_node, cache);
 364                 BUG_ON(node->data != rec);
 365                 remove_cache_extent(inode_cache, &node->cache);
 366                 free(node);
 367                 free_inode_rec(rec);
 368         }
 369 }
 370
 371 static int check_orphan_item(struct btrfs_root *root, u64 ino)
 372 {
 373         struct btrfs_path path;
 374         struct btrfs_key key;
 375         int ret;
 376
 377         key.objectid = BTRFS_ORPHAN_OBJECTID;
 378         key.type = BTRFS_ORPHAN_ITEM_KEY;
 379         key.offset = ino;
 380
 381         btrfs_init_path(&path);
 382         ret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
 383         btrfs_release_path(root, &path);
 384         if (ret > 0)
 385                 ret = -ENOENT;
 386         return ret;
 387 }
 388
 389 static int process_inode_item(struct extent_buffer *eb,
 390                               int slot, struct btrfs_key *key,
 391                               struct shared_node *active_node)
 392 {
 393         struct inode_record *rec;
 394         struct btrfs_inode_item *item;
 395
 396         rec = active_node->current;
 397         BUG_ON(rec->ino != key->objectid || rec->refs > 1);
 398         if (rec->found_inode_item) {
 399                 rec->errors |= I_ERR_DUP_INODE_ITEM;
 400                 return 1;
 401         }
 402         item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
 403         rec->nlink = btrfs_inode_nlink(eb, item);
 404         rec->isize = btrfs_inode_size(eb, item);
 405         rec->nbytes = btrfs_inode_nbytes(eb, item);
 406         rec->imode = btrfs_inode_mode(eb, item);
 407         if (btrfs_inode_flags(eb, item) & BTRFS_INODE_NODATASUM)
 408                 rec->nodatasum = 1;
 409         rec->found_inode_item = 1;
 410         if (rec->nlink == 0)
 411                 rec->errors |= I_ERR_NO_ORPHAN_ITEM;
 412         maybe_free_inode_rec(&active_node->inode_cache, rec);
 413         return 0;
 414 }
 415
 416 static struct inode_backref *get_inode_backref(struct inode_record *rec,
 417                                                 const char *name,
 418                                                 int namelen, u64 dir)
 419 {
 420         struct inode_backref *backref;
 421
 422         list_for_each_entry(backref, &rec->backrefs, list) {
 423                 if (backref->dir != dir || backref->namelen != namelen)
 424                         continue;
 425                 if (memcmp(name, backref->name, namelen))
 426                         continue;
 427                 return backref;
 428         }
 429
 430         backref = malloc(sizeof(*backref) + namelen + 1);
 431         memset(backref, 0, sizeof(*backref));
 432         backref->dir = dir;
 433         backref->namelen = namelen;
 434         memcpy(backref->name, name, namelen);
 435         backref->name[namelen] = '\0';
 436         list_add_tail(&backref->list, &rec->backrefs);
 437         return backref;
 438 }
 439
 440 static int add_inode_backref(struct cache_tree *inode_cache,
 441                              u64 ino, u64 dir, u64 index,
 442                              const char *name, int namelen,
 443                              int filetype, int itemtype, int errors)
 444 {
 445         struct inode_record *rec;
 446         struct inode_backref *backref;
 447
 448         rec = get_inode_rec(inode_cache, ino, 1);
 449         backref = get_inode_backref(rec, name, namelen, dir);
 450         if (errors)
 451                 backref->errors |= errors;
 452         if (itemtype == BTRFS_DIR_INDEX_KEY) {
 453                 if (backref->found_dir_index)
 454                         backref->errors |= REF_ERR_DUP_DIR_INDEX;
 455                 if (backref->found_inode_ref && backref->index != index)
 456                         backref->errors |= REF_ERR_INDEX_UNMATCH;
 457                 if (backref->found_dir_item && backref->filetype != filetype)
 458                         backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 459
 460                 backref->index = index;
 461                 backref->filetype = filetype;
 462                 backref->found_dir_index = 1;
 463         } else if (itemtype == BTRFS_DIR_ITEM_KEY) {
 464                 rec->found_link++;
 465                 if (backref->found_dir_item)
 466                         backref->errors |= REF_ERR_DUP_DIR_ITEM;
 467                 if (backref->found_dir_index && backref->filetype != filetype)
 468                         backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 469
 470                 backref->filetype = filetype;
 471                 backref->found_dir_item = 1;
 472         } else if (itemtype == BTRFS_INODE_REF_KEY) {
 473                 if (backref->found_inode_ref)
 474                         backref->errors |= REF_ERR_DUP_INODE_REF;
 475                 if (backref->found_dir_index && backref->index != index)
 476                         backref->errors |= REF_ERR_INDEX_UNMATCH;
 477
 478                 backref->index = index;
 479                 backref->found_inode_ref = 1;
 480         } else {
 481                 BUG_ON(1);
 482         }
 483
 484         maybe_free_inode_rec(inode_cache, rec);
 485         return 0;
 486 }
 487
 488 static int merge_inode_recs(struct inode_record *src, struct inode_record *dst,
 489                             struct cache_tree *dst_cache)
 490 {
 491         struct inode_backref *backref;
 492         u32 dir_count = 0;
 493
 494         dst->merging = 1;
 495         list_for_each_entry(backref, &src->backrefs, list) {
 496                 if (backref->found_dir_index) {
 497                         add_inode_backref(dst_cache, dst->ino, backref->dir,
 498                                         backref->index, backref->name,
 499                                         backref->namelen, backref->filetype,
 500                                         BTRFS_DIR_INDEX_KEY, backref->errors);
 501                 }
 502                 if (backref->found_dir_item) {
 503                         dir_count++;
 504                         add_inode_backref(dst_cache, dst->ino,
 505                                         backref->dir, 0, backref->name,
 506                                         backref->namelen, backref->filetype,
 507                                         BTRFS_DIR_ITEM_KEY, backref->errors);
 508                 }
 509                 if (backref->found_inode_ref) {
 510                         add_inode_backref(dst_cache, dst->ino,
 511                                         backref->dir, backref->index,
 512                                         backref->name, backref->namelen, 0,
 513                                         BTRFS_INODE_REF_KEY, backref->errors);
 514                 }
 515         }
 516
 517         if (src->found_dir_item)
 518                 dst->found_dir_item = 1;
 519         if (src->found_file_extent)
 520                 dst->found_file_extent = 1;
 521         if (src->found_csum_item)
 522                 dst->found_csum_item = 1;
 523         if (src->some_csum_missing)
 524                 dst->some_csum_missing = 1;
 525         if (dst->first_extent_gap > src->first_extent_gap)
 526                 dst->first_extent_gap = src->first_extent_gap;
 527
 528         BUG_ON(src->found_link < dir_count);
 529         dst->found_link += src->found_link - dir_count;
 530         dst->found_size += src->found_size;
 531         if (src->extent_start != (u64)-1) {
 532                 if (dst->extent_start == (u64)-1) {
 533                         dst->extent_start = src->extent_start;
 534                         dst->extent_end = src->extent_end;
 535                 } else {
 536                         if (dst->extent_end > src->extent_start)
 537                                 dst->errors |= I_ERR_FILE_EXTENT_OVERLAP;
 538                         else if (dst->extent_end < src->extent_start &&
 539                                  dst->extent_end < dst->first_extent_gap)
 540                                 dst->first_extent_gap = dst->extent_end;
 541                         if (dst->extent_end < src->extent_end)
 542                                 dst->extent_end = src->extent_end;
 543                 }
 544         }
 545
 546         dst->errors |= src->errors;
 547         if (src->found_inode_item) {
 548                 if (!dst->found_inode_item) {
 549                         dst->nlink = src->nlink;
 550                         dst->isize = src->isize;
 551                         dst->nbytes = src->nbytes;
 552                         dst->imode = src->imode;
 553                         dst->nodatasum = src->nodatasum;
 554                         dst->found_inode_item = 1;
 555                 } else {
 556                         dst->errors |= I_ERR_DUP_INODE_ITEM;
 557                 }
 558         }
 559         dst->merging = 0;
 560
 561         return 0;
 562 }
 563
 564 static int splice_shared_node(struct shared_node *src_node,
 565                               struct shared_node *dst_node)
 566 {
 567         struct cache_extent *cache;
 568         struct ptr_node *node, *ins;
 569         struct cache_tree *src, *dst;
 570         struct inode_record *rec, *conflict;
 571         u64 current_ino = 0;
 572         int splice = 0;
 573         int ret;
 574
 575         if (--src_node->refs == 0)
 576                 splice = 1;
 577         if (src_node->current)
 578                 current_ino = src_node->current->ino;
 579
 580         src = &src_node->root_cache;
 581         dst = &dst_node->root_cache;
 582 again:
 583         cache = find_first_cache_extent(src, 0);
 584         while (cache) {
 585                 node = container_of(cache, struct ptr_node, cache);
 586                 rec = node->data;
 587                 cache = next_cache_extent(cache);
 588
 589                 if (splice) {
 590                         remove_cache_extent(src, &node->cache);
 591                         ins = node;
 592                 } else {
 593                         ins = malloc(sizeof(*ins));
 594                         ins->cache.start = node->cache.start;
 595                         ins->cache.size = node->cache.size;
 596                         ins->data = rec;
 597                         rec->refs++;
 598                 }
 599                 ret = insert_existing_cache_extent(dst, &ins->cache);
 600                 if (ret == -EEXIST) {
 601                         conflict = get_inode_rec(dst, rec->ino, 1);
 602                         merge_inode_recs(rec, conflict, dst);
 603                         if (rec->checked) {
 604                                 conflict->checked = 1;
 605                                 if (dst_node->current == conflict)
 606                                         dst_node->current = NULL;
 607                         }
 608                         maybe_free_inode_rec(dst, conflict);
 609                         free_inode_rec(rec);
 610                         free(ins);
 611                 } else {
 612                         BUG_ON(ret);
 613                 }
 614         }
 615
 616         if (src == &src_node->root_cache) {
 617                 src = &src_node->inode_cache;
 618                 dst = &dst_node->inode_cache;
 619                 goto again;
 620         }
 621
 622         if (current_ino > 0 && (!dst_node->current ||
 623             current_ino > dst_node->current->ino)) {
 624                 if (dst_node->current) {
 625                         dst_node->current->checked = 1;
 626                         maybe_free_inode_rec(dst, dst_node->current);
 627                 }
 628                 dst_node->current = get_inode_rec(dst, current_ino, 1);
 629         }
 630         return 0;
 631 }
 632
 633 static void free_inode_recs(struct cache_tree *inode_cache)
 634 {
 635         struct cache_extent *cache;
 636         struct ptr_node *node;
 637         struct inode_record *rec;
 638
 639         while (1) {
 640                 cache = find_first_cache_extent(inode_cache, 0);
 641                 if (!cache)
 642                         break;
 643                 node = container_of(cache, struct ptr_node, cache);
 644                 rec = node->data;
 645                 remove_cache_extent(inode_cache, &node->cache);
 646                 free(node);
 647                 free_inode_rec(rec);
 648         }
 649 }
 650
 651 static struct shared_node *find_shared_node(struct cache_tree *shared,
 652                                             u64 bytenr)
 653 {
 654         struct cache_extent *cache;
 655         struct shared_node *node;
 656
 657         cache = find_cache_extent(shared, bytenr, 1);
 658         if (cache) {
 659                 node = container_of(cache, struct shared_node, cache);
 660                 return node;
 661         }
 662         return NULL;
 663 }
 664
 665 static int add_shared_node(struct cache_tree *shared, u64 bytenr, u32 refs)
 666 {
 667         int ret;
 668         struct shared_node *node;
 669
 670         node = calloc(1, sizeof(*node));
 671         node->cache.start = bytenr;
 672         node->cache.size = 1;
 673         cache_tree_init(&node->root_cache);
 674         cache_tree_init(&node->inode_cache);
 675         node->refs = refs;
 676
 677         ret = insert_existing_cache_extent(shared, &node->cache);
 678         BUG_ON(ret);
 679         return 0;
 680 }
 681
 682 static int enter_shared_node(struct btrfs_root *root, u64 bytenr, u32 refs,
 683                              struct walk_control *wc, int level)
 684 {
 685         struct shared_node *node;
 686         struct shared_node *dest;
 687
 688         if (level == wc->active_node)
 689                 return 0;
 690
 691         BUG_ON(wc->active_node <= level);
 692         node = find_shared_node(&wc->shared, bytenr);
 693         if (!node) {
 694                 add_shared_node(&wc->shared, bytenr, refs);
 695                 node = find_shared_node(&wc->shared, bytenr);
 696                 wc->nodes[level] = node;
 697                 wc->active_node = level;
 698                 return 0;
 699         }
 700
 701         if (wc->root_level == wc->active_node &&
 702             btrfs_root_refs(&root->root_item) == 0) {
 703                 if (--node->refs == 0) {
 704                         free_inode_recs(&node->root_cache);
 705                         free_inode_recs(&node->inode_cache);
 706                         remove_cache_extent(&wc->shared, &node->cache);
 707                         free(node);
 708                 }
 709                 return 1;
 710         }
 711
 712         dest = wc->nodes[wc->active_node];
 713         splice_shared_node(node, dest);
 714         if (node->refs == 0) {
 715                 remove_cache_extent(&wc->shared, &node->cache);
 716                 free(node);
 717         }
 718         return 1;
 719 }
 720
 721 static int leave_shared_node(struct btrfs_root *root,
 722                              struct walk_control *wc, int level)
 723 {
 724         struct shared_node *node;
 725         struct shared_node *dest;
 726         int i;
 727
 728         if (level == wc->root_level)
 729                 return 0;
 730
 731         for (i = level + 1; i < BTRFS_MAX_LEVEL; i++) {
 732                 if (wc->nodes[i])
 733                         break;
 734         }
 735         BUG_ON(i >= BTRFS_MAX_LEVEL);
 736
 737         node = wc->nodes[wc->active_node];
 738         wc->nodes[wc->active_node] = NULL;
 739         wc->active_node = i;
 740
 741         dest = wc->nodes[wc->active_node];
 742         if (wc->active_node < wc->root_level ||
 743             btrfs_root_refs(&root->root_item) > 0) {
 744                 BUG_ON(node->refs <= 1);
 745                 splice_shared_node(node, dest);
 746         } else {
 747                 BUG_ON(node->refs < 2);
 748                 node->refs--;
 749         }
 750         return 0;
 751 }
 752
 753 static int is_child_root(struct btrfs_root *root, u64 parent_root_id,
 754                          u64 child_root_id)
 755 {
 756         struct btrfs_path path;
 757         struct btrfs_key key;
 758         struct extent_buffer *leaf;
 759         int has_parent = 0;
 760         int ret;
 761
 762         btrfs_init_path(&path);
 763
 764         key.objectid = parent_root_id;
 765         key.type = BTRFS_ROOT_REF_KEY;
 766         key.offset = child_root_id;
 767         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, &path,
 768                                 0, 0);
 769         BUG_ON(ret < 0);
 770         btrfs_release_path(root, &path);
 771         if (!ret)
 772                 return 1;
 773
 774         key.objectid = child_root_id;
 775         key.type = BTRFS_ROOT_BACKREF_KEY;
 776         key.offset = 0;
 777         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, &path,
 778                                 0, 0);
 779         BUG_ON(ret <= 0);
 780
 781         while (1) {
 782                 leaf = path.nodes[0];
 783                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
 784                         ret = btrfs_next_leaf(root->fs_info->tree_root, &path);
 785                         BUG_ON(ret < 0);
 786
 787                         if (ret > 0)
 788                                 break;
 789                         leaf = path.nodes[0];
 790                 }
 791
 792                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
 793                 if (key.objectid != child_root_id ||
 794                     key.type != BTRFS_ROOT_BACKREF_KEY)
 795                         break;
 796
 797                 has_parent = 1;
 798
 799                 if (key.offset == parent_root_id) {
 800                         btrfs_release_path(root, &path);
 801                         return 1;
 802                 }
 803
 804                 path.slots[0]++;
 805         }
 806
 807         btrfs_release_path(root, &path);
 808         return has_parent? 0 : -1;
 809 }
 810
 811 static int process_dir_item(struct btrfs_root *root,
 812                             struct extent_buffer *eb,
 813                             int slot, struct btrfs_key *key,
 814                             struct shared_node *active_node)
 815 {
 816         u32 total;
 817         u32 cur = 0;
 818         u32 len;
 819         u32 name_len;
 820         u32 data_len;
 821         int error;
 822         int nritems = 0;
 823         int filetype;
 824         struct btrfs_dir_item *di;
 825         struct inode_record *rec;
 826         struct cache_tree *root_cache;
 827         struct cache_tree *inode_cache;
 828         struct btrfs_key location;
 829         char namebuf[BTRFS_NAME_LEN];
 830
 831         root_cache = &active_node->root_cache;
 832         inode_cache = &active_node->inode_cache;
 833         rec = active_node->current;
 834         rec->found_dir_item = 1;
 835
 836         di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
 837         total = btrfs_item_size_nr(eb, slot);
 838         while (cur < total) {
 839                 nritems++;
 840                 btrfs_dir_item_key_to_cpu(eb, di, &location);
 841                 name_len = btrfs_dir_name_len(eb, di);
 842                 data_len = btrfs_dir_data_len(eb, di);
 843                 filetype = btrfs_dir_type(eb, di);
 844
 845                 rec->found_size += name_len;
 846                 if (name_len <= BTRFS_NAME_LEN) {
 847                         len = name_len;
 848                         error = 0;
 849                 } else {
 850                         len = BTRFS_NAME_LEN;
 851                         error = REF_ERR_NAME_TOO_LONG;
 852                 }
 853                 read_extent_buffer(eb, namebuf, (unsigned long)(di + 1), len);
 854
 855                 if (location.type == BTRFS_INODE_ITEM_KEY) {
 856                         add_inode_backref(inode_cache, location.objectid,
 857                                           key->objectid, key->offset, namebuf,
 858                                           len, filetype, key->type, error);
 859                 } else if (location.type == BTRFS_ROOT_ITEM_KEY) {
 860                         add_inode_backref(root_cache, location.objectid,
 861                                           key->objectid, key->offset,
 862                                           namebuf, len, filetype,
 863                                           key->type, error);
 864                 } else {
 865                         fprintf(stderr, "warning line %d\n", __LINE__);
 866                 }
 867
 868                 len = sizeof(*di) + name_len + data_len;
 869                 di = (struct btrfs_dir_item *)((char *)di + len);
 870                 cur += len;
 871         }
 872         if (key->type == BTRFS_DIR_INDEX_KEY && nritems > 1)
 873                 rec->errors |= I_ERR_DUP_DIR_INDEX;
 874
 875         return 0;
 876 }
 877
 878 static int process_inode_ref(struct extent_buffer *eb,
 879                              int slot, struct btrfs_key *key,
 880                              struct shared_node *active_node)
 881 {
 882         u32 total;
 883         u32 cur = 0;
 884         u32 len;
 885         u32 name_len;
 886         u64 index;
 887         int error;
 888         struct cache_tree *inode_cache;
 889         struct btrfs_inode_ref *ref;
 890         char namebuf[BTRFS_NAME_LEN];
 891
 892         inode_cache = &active_node->inode_cache;
 893
 894         ref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
 895         total = btrfs_item_size_nr(eb, slot);
 896         while (cur < total) {
 897                 name_len = btrfs_inode_ref_name_len(eb, ref);
 898                 index = btrfs_inode_ref_index(eb, ref);
 899                 if (name_len <= BTRFS_NAME_LEN) {
 900                         len = name_len;
 901                         error = 0;
 902                 } else {
 903                         len = BTRFS_NAME_LEN;
 904                         error = REF_ERR_NAME_TOO_LONG;
 905                 }
 906                 read_extent_buffer(eb, namebuf, (unsigned long)(ref + 1), len);
 907                 add_inode_backref(inode_cache, key->objectid, key->offset,
 908                                   index, namebuf, len, 0, key->type, error);
 909
 910                 len = sizeof(*ref) + name_len;
 911                 ref = (struct btrfs_inode_ref *)((char *)ref + len);
 912                 cur += len;
 913         }
 914         return 0;
 915 }
 916
 917 static u64 count_csum_range(struct btrfs_root *root, u64 start, u64 len)
 918 {
 919         struct btrfs_key key;
 920         struct btrfs_path path;
 921         struct extent_buffer *leaf;
 922         int ret ;
 923         size_t size;
 924         u64 found = 0;
 925         u64 csum_end;
 926         u16 csum_size = btrfs_super_csum_size(&root->fs_info->super_copy);
 927
 928         btrfs_init_path(&path);
 929
 930         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
 931         key.offset = start;
 932         key.type = BTRFS_EXTENT_CSUM_KEY;
 933
 934         ret = btrfs_search_slot(NULL, root->fs_info->csum_root,
 935                                 &key, &path, 0, 0);
 936         BUG_ON(ret < 0);
 937         if (ret > 0 && path.slots[0] > 0) {
 938                 leaf = path.nodes[0];
 939                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0] - 1);
 940                 if (key.objectid == BTRFS_EXTENT_CSUM_OBJECTID &&
 941                     key.type == BTRFS_EXTENT_CSUM_KEY)
 942                         path.slots[0]--;
 943         }
 944
 945         while (len > 0) {
 946                 leaf = path.nodes[0];
 947                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
 948                         ret = btrfs_next_leaf(root->fs_info->csum_root, &path);
 949                         BUG_ON(ret < 0);
 950                         if (ret > 0)
 951                                 break;
 952                         leaf = path.nodes[0];
 953                 }
 954
 955                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
 956                 if (key.objectid != BTRFS_EXTENT_CSUM_OBJECTID ||
 957                     key.type != BTRFS_EXTENT_CSUM_KEY)
 958                         break;
 959
 960                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
 961                 if (key.offset >= start + len)
 962                         break;
 963
 964                 if (key.offset > start)
 965                         start = key.offset;
 966
 967                 size = btrfs_item_size_nr(leaf, path.slots[0]);
 968                 csum_end = key.offset + (size / csum_size) * root->sectorsize;
 969                 if (csum_end > start) {
 970                         size = min(csum_end - start, len);
 971                         len -= size;
 972                         start += size;
 973                         found += size;
 974                 }
 975
 976                 path.slots[0]++;
 977         }
 978         btrfs_release_path(root->fs_info->csum_root, &path);
 979         return found;
 980 }
 981
 982 static int process_file_extent(struct btrfs_root *root,
 983                                 struct extent_buffer *eb,
 984                                 int slot, struct btrfs_key *key,
 985                                 struct shared_node *active_node)
 986 {
 987         struct inode_record *rec;
 988         struct btrfs_file_extent_item *fi;
 989         u64 num_bytes = 0;
 990         u64 disk_bytenr = 0;
 991         u64 extent_offset = 0;
 992         u64 mask = root->sectorsize - 1;
 993         int extent_type;
 994
 995         rec = active_node->current;
 996         BUG_ON(rec->ino != key->objectid || rec->refs > 1);
 997         rec->found_file_extent = 1;
 998
 999         if (rec->extent_start == (u64)-1) {
1000                 rec->extent_start = key->offset;
1001                 rec->extent_end = key->offset;
1002         }
1003
1004         if (rec->extent_end > key->offset)
1005                 rec->errors |= I_ERR_FILE_EXTENT_OVERLAP;
1006         else if (rec->extent_end < key->offset &&
1007                  rec->extent_end < rec->first_extent_gap)
1008                 rec->first_extent_gap = rec->extent_end;
1009
1010         fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
1011         extent_type = btrfs_file_extent_type(eb, fi);
1012
1013         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1014                 num_bytes = btrfs_file_extent_inline_len(eb, fi);
1015                 if (num_bytes == 0)
1016                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1017                 rec->found_size += num_bytes;
1018                 num_bytes = (num_bytes + mask) & ~mask;
1019         } else if (extent_type == BTRFS_FILE_EXTENT_REG ||
1020                    extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1021                 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1022                 disk_bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1023                 extent_offset = btrfs_file_extent_offset(eb, fi);
1024                 if (num_bytes == 0 || (num_bytes & mask))
1025                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1026                 if (num_bytes + extent_offset >
1027                     btrfs_file_extent_ram_bytes(eb, fi))
1028                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1029                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC &&
1030                     (btrfs_file_extent_compression(eb, fi) ||
1031                      btrfs_file_extent_encryption(eb, fi) ||
1032                      btrfs_file_extent_other_encoding(eb, fi)))
1033                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1034                 if (disk_bytenr > 0)
1035                         rec->found_size += num_bytes;
1036         } else {
1037                 rec->errors |= I_ERR_BAD_FILE_EXTENT;
1038         }
1039         rec->extent_end = key->offset + num_bytes;
1040
1041         if (disk_bytenr > 0) {
1042                 u64 found;
1043                 if (btrfs_file_extent_compression(eb, fi))
1044                         num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1045                 else
1046                         disk_bytenr += extent_offset;
1047
1048                 found = count_csum_range(root, disk_bytenr, num_bytes);
1049                 if (extent_type == BTRFS_FILE_EXTENT_REG) {
1050                         if (found > 0)
1051                                 rec->found_csum_item = 1;
1052                         if (found < num_bytes)
1053                                 rec->some_csum_missing = 1;
1054                 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1055                         if (found > 0)
1056                                 rec->errors |= I_ERR_ODD_CSUM_ITEM;
1057                 }
1058         }
1059         return 0;
1060 }
1061
1062 static int process_one_leaf(struct btrfs_root *root, struct extent_buffer *eb,
1063                             struct walk_control *wc)
1064 {
1065         struct btrfs_key key;
1066         u32 nritems;
1067         int i;
1068         int ret = 0;
1069         struct cache_tree *inode_cache;
1070         struct shared_node *active_node;
1071
1072         if (wc->root_level == wc->active_node &&
1073             btrfs_root_refs(&root->root_item) == 0)
1074                 return 0;
1075
1076         active_node = wc->nodes[wc->active_node];
1077         inode_cache = &active_node->inode_cache;
1078         nritems = btrfs_header_nritems(eb);
1079         for (i = 0; i < nritems; i++) {
1080                 btrfs_item_key_to_cpu(eb, &key, i);
1081
1082                 if (key.objectid == BTRFS_FREE_SPACE_OBJECTID)
1083                         continue;
1084
1085                 if (active_node->current == NULL ||
1086                     active_node->current->ino < key.objectid) {
1087                         if (active_node->current) {
1088                                 active_node->current->checked = 1;
1089                                 maybe_free_inode_rec(inode_cache,
1090                                                      active_node->current);
1091                         }
1092                         active_node->current = get_inode_rec(inode_cache,
1093                                                              key.objectid, 1);
1094                 }
1095                 switch (key.type) {
1096                 case BTRFS_DIR_ITEM_KEY:
1097                 case BTRFS_DIR_INDEX_KEY:
1098                         ret = process_dir_item(root, eb, i, &key, active_node);
1099                         break;
1100                 case BTRFS_INODE_REF_KEY:
1101                         ret = process_inode_ref(eb, i, &key, active_node);
1102                         break;
1103                 case BTRFS_INODE_ITEM_KEY:
1104                         ret = process_inode_item(eb, i, &key, active_node);
1105                         break;
1106                 case BTRFS_EXTENT_DATA_KEY:
1107                         ret = process_file_extent(root, eb, i, &key,
1108                                                   active_node);
1109                         break;
1110                 default:
1111                         break;
1112                 };
1113         }
1114         return ret;
1115 }
1116
1117 static void reada_walk_down(struct btrfs_root *root,
1118                             struct extent_buffer *node, int slot)
1119 {
1120         u64 bytenr;
1121         u64 ptr_gen;
1122         u32 nritems;
1123         u32 blocksize;
1124         int i;
1125         int ret;
1126         int level;
1127
1128         level = btrfs_header_level(node);
1129         if (level != 1)
1130                 return;
1131
1132         nritems = btrfs_header_nritems(node);
1133         blocksize = btrfs_level_size(root, level - 1);
1134         for (i = slot; i < nritems; i++) {
1135                 bytenr = btrfs_node_blockptr(node, i);
1136                 ptr_gen = btrfs_node_ptr_generation(node, i);
1137                 ret = readahead_tree_block(root, bytenr, blocksize, ptr_gen);
1138                 if (ret)
1139                         break;
1140         }
1141 }
1142
1143 static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
1144                           struct walk_control *wc, int *level)
1145 {
1146         u64 bytenr;
1147         u64 ptr_gen;
1148         struct extent_buffer *next;
1149         struct extent_buffer *cur;
1150         u32 blocksize;
1151         int ret;
1152         u64 refs;
1153
1154         WARN_ON(*level < 0);
1155         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1156         ret = btrfs_lookup_extent_info(NULL, root,
1157                                        path->nodes[*level]->start,
1158                                        path->nodes[*level]->len, &refs, NULL);
1159         if (ret < 0)
1160                 goto out;
1161
1162         if (refs > 1) {
1163                 ret = enter_shared_node(root, path->nodes[*level]->start,
1164                                         refs, wc, *level);
1165                 if (ret > 0)
1166                         goto out;
1167         }
1168
1169         while (*level >= 0) {
1170                 WARN_ON(*level < 0);
1171                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1172                 cur = path->nodes[*level];
1173
1174                 if (btrfs_header_level(cur) != *level)
1175                         WARN_ON(1);
1176
1177                 if (path->slots[*level] >= btrfs_header_nritems(cur))
1178                         break;
1179                 if (*level == 0) {
1180                         ret = process_one_leaf(root, cur, wc);
1181                         break;
1182                 }
1183                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1184                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1185                 blocksize = btrfs_level_size(root, *level - 1);
1186                 ret = btrfs_lookup_extent_info(NULL, root, bytenr, blocksize,
1187                                                &refs, NULL);
1188                 if (ret < 0)
1189                         refs = 0;
1190
1191                 if (refs > 1) {
1192                         ret = enter_shared_node(root, bytenr, refs,
1193                                                 wc, *level - 1);
1194                         if (ret > 0) {
1195                                 path->slots[*level]++;
1196                                 continue;
1197                         }
1198                 }
1199
1200                 next = btrfs_find_tree_block(root, bytenr, blocksize);
1201                 if (!next || !btrfs_buffer_uptodate(next, ptr_gen)) {
1202                         free_extent_buffer(next);
1203                         reada_walk_down(root, cur, path->slots[*level]);
1204                         next = read_tree_block(root, bytenr, blocksize,
1205                                                ptr_gen);
1206                 }
1207
1208                 *level = *level - 1;
1209                 free_extent_buffer(path->nodes[*level]);
1210                 path->nodes[*level] = next;
1211                 path->slots[*level] = 0;
1212         }
1213 out:
1214         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
1215         return 0;
1216 }
1217
1218 static int walk_up_tree(struct btrfs_root *root, struct btrfs_path *path,
1219                         struct walk_control *wc, int *level)
1220 {
1221         int i;
1222         struct extent_buffer *leaf;
1223
1224         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
1225                 leaf = path->nodes[i];
1226                 if (path->slots[i] + 1 < btrfs_header_nritems(leaf)) {
1227                         path->slots[i]++;
1228                         *level = i;
1229                         return 0;
1230                 } else {
1231                         free_extent_buffer(path->nodes[*level]);
1232                         path->nodes[*level] = NULL;
1233                         BUG_ON(*level > wc->active_node);
1234                         if (*level == wc->active_node)
1235                                 leave_shared_node(root, wc, *level);
1236                         *level = i + 1;
1237                 }
1238         }
1239         return 1;
1240 }
1241
1242 static int check_root_dir(struct inode_record *rec)
1243 {
1244         struct inode_backref *backref;
1245         int ret = -1;
1246
1247         if (!rec->found_inode_item || rec->errors)
1248                 goto out;
1249         if (rec->nlink != 1 || rec->found_link != 0)
1250                 goto out;
1251         if (list_empty(&rec->backrefs))
1252                 goto out;
1253         backref = list_entry(rec->backrefs.next, struct inode_backref, list);
1254         if (!backref->found_inode_ref)
1255                 goto out;
1256         if (backref->index != 0 || backref->namelen != 2 ||
1257             memcmp(backref->name, "..", 2))
1258                 goto out;
1259         if (backref->found_dir_index || backref->found_dir_item)
1260                 goto out;
1261         ret = 0;
1262 out:
1263         return ret;
1264 }
1265
1266 static int check_inode_recs(struct btrfs_root *root,
1267                             struct cache_tree *inode_cache)
1268 {
1269         struct cache_extent *cache;
1270         struct ptr_node *node;
1271         struct inode_record *rec;
1272         struct inode_backref *backref;
1273         int ret;
1274         u64 error = 0;
1275         u64 root_dirid = btrfs_root_dirid(&root->root_item);
1276
1277         if (btrfs_root_refs(&root->root_item) == 0) {
1278                 if (!cache_tree_empty(inode_cache))
1279                         fprintf(stderr, "warning line %d\n", __LINE__);
1280                 return 0;
1281         }
1282
1283         rec = get_inode_rec(inode_cache, root_dirid, 0);
1284         if (rec) {
1285                 ret = check_root_dir(rec);
1286                 if (ret) {
1287                         fprintf(stderr, "root %llu root dir %llu error\n",
1288                                 (unsigned long long)root->root_key.objectid,
1289                                 (unsigned long long)root_dirid);
1290                         error++;
1291                 }
1292         } else {
1293                 fprintf(stderr, "root %llu root dir %llu not found\n",
1294                         (unsigned long long)root->root_key.objectid,
1295                         (unsigned long long)root_dirid);
1296         }
1297
1298         while (1) {
1299                 cache = find_first_cache_extent(inode_cache, 0);
1300                 if (!cache)
1301                         break;
1302                 node = container_of(cache, struct ptr_node, cache);
1303                 rec = node->data;
1304                 remove_cache_extent(inode_cache, &node->cache);
1305                 free(node);
1306                 if (rec->ino == root_dirid ||
1307                     rec->ino == BTRFS_ORPHAN_OBJECTID) {
1308                         free_inode_rec(rec);
1309                         continue;
1310                 }
1311
1312                 if (rec->errors & I_ERR_NO_ORPHAN_ITEM) {
1313                         ret = check_orphan_item(root, rec->ino);
1314                         if (ret == 0)
1315                                 rec->errors &= ~I_ERR_NO_ORPHAN_ITEM;
1316                         if (can_free_inode_rec(rec)) {
1317                                 free_inode_rec(rec);
1318                                 continue;
1319                         }
1320                 }
1321
1322                 error++;
1323                 if (!rec->found_inode_item)
1324                         rec->errors |= I_ERR_NO_INODE_ITEM;
1325                 if (rec->found_link != rec->nlink)
1326                         rec->errors |= I_ERR_LINK_COUNT_WRONG;
1327                 fprintf(stderr, "root %llu inode %llu errors %x\n",
1328                         (unsigned long long) root->root_key.objectid,
1329                         (unsigned long long) rec->ino, rec->errors);
1330                 list_for_each_entry(backref, &rec->backrefs, list) {
1331                         if (!backref->found_dir_item)
1332                                 backref->errors |= REF_ERR_NO_DIR_ITEM;
1333                         if (!backref->found_dir_index)
1334                                 backref->errors |= REF_ERR_NO_DIR_INDEX;
1335                         if (!backref->found_inode_ref)
1336                                 backref->errors |= REF_ERR_NO_INODE_REF;
1337                         fprintf(stderr, "\tunresolved ref dir %llu index %llu"
1338                                 " namelen %u name %s filetype %d error %x\n",
1339                                 (unsigned long long)backref->dir,
1340                                 (unsigned long long)backref->index,
1341                                 backref->namelen, backref->name,
1342                                 backref->filetype, backref->errors);
1343                 }
1344                 free_inode_rec(rec);
1345         }
1346         return (error > 0) ? -1 : 0;
1347 }
1348
1349 static struct root_record *get_root_rec(struct cache_tree *root_cache,
1350                                         u64 objectid)
1351 {
1352         struct cache_extent *cache;
1353         struct root_record *rec = NULL;
1354         int ret;
1355
1356         cache = find_cache_extent(root_cache, objectid, 1);
1357         if (cache) {
1358                 rec = container_of(cache, struct root_record, cache);
1359         } else {
1360                 rec = calloc(1, sizeof(*rec));
1361                 rec->objectid = objectid;
1362                 INIT_LIST_HEAD(&rec->backrefs);
1363                 rec->cache.start = objectid;
1364                 rec->cache.size = 1;
1365
1366                 ret = insert_existing_cache_extent(root_cache, &rec->cache);
1367                 BUG_ON(ret);
1368         }
1369         return rec;
1370 }
1371
1372 static struct root_backref *get_root_backref(struct root_record *rec,
1373                                              u64 ref_root, u64 dir, u64 index,
1374                                              const char *name, int namelen)
1375 {
1376         struct root_backref *backref;
1377
1378         list_for_each_entry(backref, &rec->backrefs, list) {
1379                 if (backref->ref_root != ref_root || backref->dir != dir ||
1380                     backref->namelen != namelen)
1381                         continue;
1382                 if (memcmp(name, backref->name, namelen))
1383                         continue;
1384                 return backref;
1385         }
1386
1387         backref = malloc(sizeof(*backref) + namelen + 1);
1388         memset(backref, 0, sizeof(*backref));
1389         backref->ref_root = ref_root;
1390         backref->dir = dir;
1391         backref->index = index;
1392         backref->namelen = namelen;
1393         memcpy(backref->name, name, namelen);
1394         backref->name[namelen] = '\0';
1395         list_add_tail(&backref->list, &rec->backrefs);
1396         return backref;
1397 }
1398
1399 static void free_root_recs(struct cache_tree *root_cache)
1400 {
1401         struct cache_extent *cache;
1402         struct root_record *rec;
1403         struct root_backref *backref;
1404
1405         while (1) {
1406                 cache = find_first_cache_extent(root_cache, 0);
1407                 if (!cache)
1408                         break;
1409                 rec = container_of(cache, struct root_record, cache);
1410                 remove_cache_extent(root_cache, &rec->cache);
1411
1412                 while (!list_empty(&rec->backrefs)) {
1413                         backref = list_entry(rec->backrefs.next,
1414                                              struct root_backref, list);
1415                         list_del(&backref->list);
1416                         free(backref);
1417                 }
1418                 kfree(rec);
1419         }
1420 }
1421
1422 static int add_root_backref(struct cache_tree *root_cache,
1423                             u64 root_id, u64 ref_root, u64 dir, u64 index,
1424                             const char *name, int namelen,
1425                             int item_type, int errors)
1426 {
1427         struct root_record *rec;
1428         struct root_backref *backref;
1429
1430         rec = get_root_rec(root_cache, root_id);
1431         backref = get_root_backref(rec, ref_root, dir, index, name, namelen);
1432
1433         backref->errors |= errors;
1434
1435         if (item_type != BTRFS_DIR_ITEM_KEY) {
1436                 if (backref->found_dir_index || backref->found_back_ref ||
1437                     backref->found_forward_ref) {
1438                         if (backref->index != index)
1439                                 backref->errors |= REF_ERR_INDEX_UNMATCH;
1440                 } else {
1441                         backref->index = index;
1442                 }
1443         }
1444
1445         if (item_type == BTRFS_DIR_ITEM_KEY) {
1446                 backref->found_dir_item = 1;
1447                 backref->reachable = 1;
1448                 rec->found_ref++;
1449         } else if (item_type == BTRFS_DIR_INDEX_KEY) {
1450                 backref->found_dir_index = 1;
1451         } else if (item_type == BTRFS_ROOT_REF_KEY) {
1452                 if (backref->found_forward_ref)
1453                         backref->errors |= REF_ERR_DUP_ROOT_REF;
1454                 backref->found_forward_ref = 1;
1455         } else if (item_type == BTRFS_ROOT_BACKREF_KEY) {
1456                 if (backref->found_back_ref)
1457                         backref->errors |= REF_ERR_DUP_ROOT_BACKREF;
1458                 backref->found_back_ref = 1;
1459         } else {
1460                 BUG_ON(1);
1461         }
1462
1463         return 0;
1464 }
1465
1466 static int merge_root_recs(struct btrfs_root *root,
1467                            struct cache_tree *src_cache,
1468                            struct cache_tree *dst_cache)
1469 {
1470         struct cache_extent *cache;
1471         struct ptr_node *node;
1472         struct inode_record *rec;
1473         struct inode_backref *backref;
1474
1475         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
1476                 free_inode_recs(src_cache);
1477                 return 0;
1478         }
1479
1480         while (1) {
1481                 cache = find_first_cache_extent(src_cache, 0);
1482                 if (!cache)
1483                         break;
1484                 node = container_of(cache, struct ptr_node, cache);
1485                 rec = node->data;
1486                 remove_cache_extent(src_cache, &node->cache);
1487                 free(node);
1488
1489                 if (!is_child_root(root, root->objectid, rec->ino))
1490                         goto skip;
1491
1492                 list_for_each_entry(backref, &rec->backrefs, list) {
1493                         BUG_ON(backref->found_inode_ref);
1494                         if (backref->found_dir_item)
1495                                 add_root_backref(dst_cache, rec->ino,
1496                                         root->root_key.objectid, backref->dir,
1497                                         backref->index, backref->name,
1498                                         backref->namelen, BTRFS_DIR_ITEM_KEY,
1499                                         backref->errors);
1500                         if (backref->found_dir_index)
1501                                 add_root_backref(dst_cache, rec->ino,
1502                                         root->root_key.objectid, backref->dir,
1503                                         backref->index, backref->name,
1504                                         backref->namelen, BTRFS_DIR_INDEX_KEY,
1505                                         backref->errors);
1506                 }
1507 skip:
1508                 free_inode_rec(rec);
1509         }
1510         return 0;
1511 }
1512
1513 static int check_root_refs(struct btrfs_root *root,
1514                            struct cache_tree *root_cache)
1515 {
1516         struct root_record *rec;
1517         struct root_record *ref_root;
1518         struct root_backref *backref;
1519         struct cache_extent *cache;
1520         int loop = 1;
1521         int ret;
1522         int error;
1523         int errors = 0;
1524
1525         rec = get_root_rec(root_cache, BTRFS_FS_TREE_OBJECTID);
1526         rec->found_ref = 1;
1527
1528         /* fixme: this can not detect circular references */
1529         while (loop) {
1530                 loop = 0;
1531                 cache = find_first_cache_extent(root_cache, 0);
1532                 while (1) {
1533                         if (!cache)
1534                                 break;
1535                         rec = container_of(cache, struct root_record, cache);
1536                         cache = next_cache_extent(cache);
1537
1538                         if (rec->found_ref == 0)
1539                                 continue;
1540
1541                         list_for_each_entry(backref, &rec->backrefs, list) {
1542                                 if (!backref->reachable)
1543                                         continue;
1544
1545                                 ref_root = get_root_rec(root_cache,
1546                                                         backref->ref_root);
1547                                 if (ref_root->found_ref > 0)
1548                                         continue;
1549
1550                                 backref->reachable = 0;
1551                                 rec->found_ref--;
1552                                 if (rec->found_ref == 0)
1553                                         loop = 1;
1554                         }
1555                 }
1556         }
1557
1558         cache = find_first_cache_extent(root_cache, 0);
1559         while (1) {
1560                 if (!cache)
1561                         break;
1562                 rec = container_of(cache, struct root_record, cache);
1563                 cache = next_cache_extent(cache);
1564
1565                 if (rec->found_ref == 0 &&
1566                     rec->objectid >= BTRFS_FIRST_FREE_OBJECTID &&
1567                     rec->objectid <= BTRFS_LAST_FREE_OBJECTID) {
1568                         ret = check_orphan_item(root->fs_info->tree_root,
1569                                                 rec->objectid);
1570                         if (ret == 0)
1571                                 continue;
1572                         errors++;
1573                         fprintf(stderr, "fs tree %llu not referenced\n",
1574                                 (unsigned long long)rec->objectid);
1575                 }
1576
1577                 error = 0;
1578                 if (rec->found_ref > 0 && !rec->found_root_item)
1579                         error = 1;
1580                 list_for_each_entry(backref, &rec->backrefs, list) {
1581                         if (!backref->found_dir_item)
1582                                 backref->errors |= REF_ERR_NO_DIR_ITEM;
1583                         if (!backref->found_dir_index)
1584                                 backref->errors |= REF_ERR_NO_DIR_INDEX;
1585                         if (!backref->found_back_ref)
1586                                 backref->errors |= REF_ERR_NO_ROOT_BACKREF;
1587                         if (!backref->found_forward_ref)
1588                                 backref->errors |= REF_ERR_NO_ROOT_REF;
1589                         if (backref->reachable && backref->errors)
1590                                 error = 1;
1591                 }
1592                 if (!error)
1593                         continue;
1594
1595                 errors++;
1596                 fprintf(stderr, "fs tree %llu refs %u %s\n",
1597                         (unsigned long long)rec->objectid, rec->found_ref,
1598                          rec->found_root_item ? "" : "not found");
1599
1600                 list_for_each_entry(backref, &rec->backrefs, list) {
1601                         if (!backref->reachable)
1602                                 continue;
1603                         if (!backref->errors && rec->found_root_item)
1604                                 continue;
1605                         fprintf(stderr, "\tunresolved ref root %llu dir %llu"
1606                                 " index %llu namelen %u name %s error %x\n",
1607                                 (unsigned long long)backref->ref_root,
1608                                 (unsigned long long)backref->dir,
1609                                 (unsigned long long)backref->index,
1610                                 backref->namelen, backref->name,
1611                                 backref->errors);
1612                 }
1613         }
1614         return errors > 0 ? 1 : 0;
1615 }
1616
1617 static int process_root_ref(struct extent_buffer *eb, int slot,
1618                             struct btrfs_key *key,
1619                             struct cache_tree *root_cache)
1620 {
1621         u64 dirid;
1622         u64 index;
1623         u32 len;
1624         u32 name_len;
1625         struct btrfs_root_ref *ref;
1626         char namebuf[BTRFS_NAME_LEN];
1627         int error;
1628
1629         ref = btrfs_item_ptr(eb, slot, struct btrfs_root_ref);
1630
1631         dirid = btrfs_root_ref_dirid(eb, ref);
1632         index = btrfs_root_ref_sequence(eb, ref);
1633         name_len = btrfs_root_ref_name_len(eb, ref);
1634
1635         if (name_len <= BTRFS_NAME_LEN) {
1636                 len = name_len;
1637                 error = 0;
1638         } else {
1639                 len = BTRFS_NAME_LEN;
1640                 error = REF_ERR_NAME_TOO_LONG;
1641         }
1642         read_extent_buffer(eb, namebuf, (unsigned long)(ref + 1), len);
1643
1644         if (key->type == BTRFS_ROOT_REF_KEY) {
1645                 add_root_backref(root_cache, key->offset, key->objectid, dirid,
1646                                  index, namebuf, len, key->type, error);
1647         } else {
1648                 add_root_backref(root_cache, key->objectid, key->offset, dirid,
1649                                  index, namebuf, len, key->type, error);
1650         }
1651         return 0;
1652 }
1653
1654 static int check_fs_root(struct btrfs_root *root,
1655                          struct cache_tree *root_cache,
1656                          struct walk_control *wc)
1657 {
1658         int ret = 0;
1659         int wret;
1660         int level;
1661         struct btrfs_path path;
1662         struct shared_node root_node;
1663         struct root_record *rec;
1664         struct btrfs_root_item *root_item = &root->root_item;
1665
1666         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1667                 rec = get_root_rec(root_cache, root->root_key.objectid);
1668                 if (btrfs_root_refs(root_item) > 0)
1669                         rec->found_root_item = 1;
1670         }
1671
1672         btrfs_init_path(&path);
1673         memset(&root_node, 0, sizeof(root_node));
1674         cache_tree_init(&root_node.root_cache);
1675         cache_tree_init(&root_node.inode_cache);
1676
1677         level = btrfs_header_level(root->node);
1678         memset(wc->nodes, 0, sizeof(wc->nodes));
1679         wc->nodes[level] = &root_node;
1680         wc->active_node = level;
1681         wc->root_level = level;
1682
1683         if (btrfs_root_refs(root_item) > 0 ||
1684             btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1685                 path.nodes[level] = root->node;
1686                 extent_buffer_get(root->node);
1687                 path.slots[level] = 0;
1688         } else {
1689                 struct btrfs_key key;
1690                 struct btrfs_disk_key found_key;
1691
1692                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1693                 level = root_item->drop_level;
1694                 path.lowest_level = level;
1695                 wret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
1696                 BUG_ON(wret < 0);
1697                 btrfs_node_key(path.nodes[level], &found_key,
1698                                 path.slots[level]);
1699                 WARN_ON(memcmp(&found_key, &root_item->drop_progress,
1700                                         sizeof(found_key)));
1701         }
1702
1703         while (1) {
1704                 wret = walk_down_tree(root, &path, wc, &level);
1705                 if (wret < 0)
1706                         ret = wret;
1707                 if (wret != 0)
1708                         break;
1709
1710                 wret = walk_up_tree(root, &path, wc, &level);
1711                 if (wret < 0)
1712                         ret = wret;
1713                 if (wret != 0)
1714                         break;
1715         }
1716         btrfs_release_path(root, &path);
1717
1718         merge_root_recs(root, &root_node.root_cache, root_cache);
1719
1720         if (root_node.current) {
1721                 root_node.current->checked = 1;
1722                 maybe_free_inode_rec(&root_node.inode_cache,
1723                                 root_node.current);
1724         }
1725
1726         ret = check_inode_recs(root, &root_node.inode_cache);
1727         return ret;
1728 }
1729
1730 static int fs_root_objectid(u64 objectid)
1731 {
1732         if (objectid == BTRFS_FS_TREE_OBJECTID ||
1733             objectid == BTRFS_TREE_RELOC_OBJECTID ||
1734             objectid == BTRFS_DATA_RELOC_TREE_OBJECTID ||
1735             (objectid >= BTRFS_FIRST_FREE_OBJECTID &&
1736              objectid <= BTRFS_LAST_FREE_OBJECTID))
1737                 return 1;
1738         return 0;
1739 }
1740
1741 static int check_fs_roots(struct btrfs_root *root,
1742                           struct cache_tree *root_cache)
1743 {
1744         struct btrfs_path path;
1745         struct btrfs_key key;
1746         struct walk_control wc;
1747         struct extent_buffer *leaf;
1748         struct btrfs_root *tmp_root;
1749         struct btrfs_root *tree_root = root->fs_info->tree_root;
1750         int ret;
1751         int err = 0;
1752
1753         memset(&wc, 0, sizeof(wc));
1754         cache_tree_init(&wc.shared);
1755         btrfs_init_path(&path);
1756
1757         key.offset = 0;
1758         key.objectid = 0;
1759         key.type = BTRFS_ROOT_ITEM_KEY;
1760         ret = btrfs_search_slot(NULL, tree_root, &key, &path, 0, 0);
1761         BUG_ON(ret < 0);
1762         while (1) {
1763                 leaf = path.nodes[0];
1764                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1765                         ret = btrfs_next_leaf(tree_root, &path);
1766                         if (ret != 0)
1767                                 break;
1768                         leaf = path.nodes[0];
1769                 }
1770                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1771                 if (key.type == BTRFS_ROOT_ITEM_KEY &&
1772                     fs_root_objectid(key.objectid)) {
1773                         tmp_root = btrfs_read_fs_root_no_cache(root->fs_info,
1774                                                                &key);
1775                         ret = check_fs_root(tmp_root, root_cache, &wc);
1776                         if (ret)
1777                                 err = 1;
1778                         btrfs_free_fs_root(root->fs_info, tmp_root);
1779                 } else if (key.type == BTRFS_ROOT_REF_KEY ||
1780                            key.type == BTRFS_ROOT_BACKREF_KEY) {
1781                         process_root_ref(leaf, path.slots[0], &key,
1782                                          root_cache);
1783                 }
1784                 path.slots[0]++;
1785         }
1786         btrfs_release_path(tree_root, &path);
1787
1788         if (!cache_tree_empty(&wc.shared))
1789                 fprintf(stderr, "warning line %d\n", __LINE__);
1790
1791         return err;
1792 }
1793
1794 static int all_backpointers_checked(struct extent_record *rec, int print_errs)
1795 {
1796         struct list_head *cur = rec->backrefs.next;
1797         struct extent_backref *back;
1798         struct tree_backref *tback;
1799         struct data_backref *dback;
1800         u64 found = 0;
1801         int err = 0;
1802
1803         while(cur != &rec->backrefs) {
1804                 back = list_entry(cur, struct extent_backref, list);
1805                 cur = cur->next;
1806                 if (!back->found_extent_tree) {
1807                         err = 1;
1808                         if (!print_errs)
1809                                 goto out;
1810                         if (back->is_data) {
1811                                 dback = (struct data_backref *)back;
1812                                 fprintf(stderr, "Backref %llu %s %llu"
1813                                         " owner %llu offset %llu num_refs %lu"
1814                                         " not found in extent tree\n",
1815                                         (unsigned long long)rec->start,
1816                                         back->full_backref ?
1817                                         "parent" : "root",
1818                                         back->full_backref ?
1819                                         (unsigned long long)dback->parent:
1820                                         (unsigned long long)dback->root,
1821                                         (unsigned long long)dback->owner,
1822                                         (unsigned long long)dback->offset,
1823                                         (unsigned long)dback->num_refs);
1824                         } else {
1825                                 tback = (struct tree_backref *)back;
1826                                 fprintf(stderr, "Backref %llu parent %llu"
1827                                         " root %llu not found in extent tree\n",
1828                                         (unsigned long long)rec->start,
1829                                         (unsigned long long)tback->parent,
1830                                         (unsigned long long)tback->root);
1831                         }
1832                 }
1833                 if (!back->is_data && !back->found_ref) {
1834                         err = 1;
1835                         if (!print_errs)
1836                                 goto out;
1837                         tback = (struct tree_backref *)back;
1838                         fprintf(stderr, "Backref %llu %s %llu not referenced back %p\n",
1839                                 (unsigned long long)rec->start,
1840                                 back->full_backref ? "parent" : "root",
1841                                 back->full_backref ?
1842                                 (unsigned long long)tback->parent :
1843                                 (unsigned long long)tback->root, back);
1844                 }
1845                 if (back->is_data) {
1846                         dback = (struct data_backref *)back;
1847                         if (dback->found_ref != dback->num_refs) {
1848                                 err = 1;
1849                                 if (!print_errs)
1850                                         goto out;
1851                                 fprintf(stderr, "Incorrect local backref count"
1852                                         " on %llu %s %llu owner %llu"
1853                                         " offset %llu found %u wanted %u back %p\n",
1854                                         (unsigned long long)rec->start,
1855                                         back->full_backref ?
1856                                         "parent" : "root",
1857                                         back->full_backref ?
1858                                         (unsigned long long)dback->parent:
1859                                         (unsigned long long)dback->root,
1860                                         (unsigned long long)dback->owner,
1861                                         (unsigned long long)dback->offset,
1862                                         dback->found_ref, dback->num_refs, back);
1863                         }
1864                 }
1865                 if (!back->is_data) {
1866                         found += 1;
1867                 } else {
1868                         dback = (struct data_backref *)back;
1869                         found += dback->found_ref;
1870                 }
1871         }
1872         if (found != rec->refs) {
1873                 err = 1;
1874                 if (!print_errs)
1875                         goto out;
1876                 fprintf(stderr, "Incorrect global backref count "
1877                         "on %llu found %llu wanted %llu\n",
1878                         (unsigned long long)rec->start,
1879                         (unsigned long long)found,
1880                         (unsigned long long)rec->refs);
1881         }
1882 out:
1883         return err;
1884 }
1885
1886 static int free_all_extent_backrefs(struct extent_record *rec)
1887 {
1888         struct extent_backref *back;
1889         struct list_head *cur;
1890         while (!list_empty(&rec->backrefs)) {
1891                 cur = rec->backrefs.next;
1892                 back = list_entry(cur, struct extent_backref, list);
1893                 list_del(cur);
1894                 free(back);
1895         }
1896         return 0;
1897 }
1898
1899 static int maybe_free_extent_rec(struct cache_tree *extent_cache,
1900                                  struct extent_record *rec)
1901 {
1902         if (rec->content_checked && rec->owner_ref_checked &&
1903             rec->extent_item_refs == rec->refs && rec->refs > 0 &&
1904             !all_backpointers_checked(rec, 0)) {
1905                 remove_cache_extent(extent_cache, &rec->cache);
1906                 free_all_extent_backrefs(rec);
1907                 free(rec);
1908         }
1909         return 0;
1910 }
1911
1912 static int check_owner_ref(struct btrfs_root *root,
1913                             struct extent_record *rec,
1914                             struct extent_buffer *buf)
1915 {
1916         struct extent_backref *node;
1917         struct tree_backref *back;
1918         struct btrfs_root *ref_root;
1919         struct btrfs_key key;
1920         struct btrfs_path path;
1921         int level;
1922         int found = 0;
1923
1924         list_for_each_entry(node, &rec->backrefs, list) {
1925                 if (node->is_data)
1926                         continue;
1927                 if (!node->found_ref)
1928                         continue;
1929                 if (node->full_backref)
1930                         continue;
1931                 back = (struct tree_backref *)node;
1932                 if (btrfs_header_owner(buf) == back->root)
1933                         return 0;
1934         }
1935         BUG_ON(rec->is_root);
1936
1937         /* try to find the block by search corresponding fs tree */
1938         key.objectid = btrfs_header_owner(buf);
1939         key.type = BTRFS_ROOT_ITEM_KEY;
1940         key.offset = (u64)-1;
1941
1942         ref_root = btrfs_read_fs_root(root->fs_info, &key);
1943         BUG_ON(IS_ERR(ref_root));
1944
1945         level = btrfs_header_level(buf);
1946         if (level == 0)
1947                 btrfs_item_key_to_cpu(buf, &key, 0);
1948         else
1949                 btrfs_node_key_to_cpu(buf, &key, 0);
1950
1951         btrfs_init_path(&path);
1952         path.lowest_level = level + 1;
1953         btrfs_search_slot(NULL, ref_root, &key, &path, 0, 0);
1954
1955         if (buf->start == btrfs_node_blockptr(path.nodes[level + 1],
1956                                               path.slots[level + 1]))
1957                 found = 1;
1958
1959         btrfs_release_path(ref_root, &path);
1960         return found ? 0 : 1;
1961 }
1962
1963 static int is_extent_tree_record(struct extent_record *rec)
1964 {
1965         struct list_head *cur = rec->backrefs.next;
1966         struct extent_backref *node;
1967         struct tree_backref *back;
1968         int is_extent = 0;
1969
1970         while(cur != &rec->backrefs) {
1971                 node = list_entry(cur, struct extent_backref, list);
1972                 cur = cur->next;
1973                 if (node->is_data)
1974                         return 0;
1975                 back = (struct tree_backref *)node;
1976                 if (node->full_backref)
1977                         return 0;
1978                 if (back->root == BTRFS_EXTENT_TREE_OBJECTID)
1979                         is_extent = 1;
1980         }
1981         return is_extent;
1982 }
1983
1984
1985 static int record_bad_block_io(struct btrfs_fs_info *info,
1986                                struct cache_tree *extent_cache,
1987                                u64 start, u64 len)
1988 {
1989         struct extent_record *rec;
1990         struct cache_extent *cache;
1991         struct btrfs_key key;
1992
1993         cache = find_cache_extent(extent_cache, start, len);
1994         if (!cache)
1995                 return 0;
1996
1997         rec = container_of(cache, struct extent_record, cache);
1998         if (!is_extent_tree_record(rec))
1999                 return 0;
2000
2001         btrfs_disk_key_to_cpu(&key, &rec->parent_key);
2002         return btrfs_add_corrupt_extent_record(info, &key, start, len, 0);
2003 }
2004
2005 static int check_block(struct btrfs_root *root,
2006                        struct cache_tree *extent_cache,
2007                        struct extent_buffer *buf, u64 flags)
2008 {
2009         struct extent_record *rec;
2010         struct cache_extent *cache;
2011         struct btrfs_key key;
2012         int ret = 1;
2013         int level;
2014
2015         cache = find_cache_extent(extent_cache, buf->start, buf->len);
2016         if (!cache)
2017                 return 1;
2018         rec = container_of(cache, struct extent_record, cache);
2019         rec->generation = btrfs_header_generation(buf);
2020
2021         level = btrfs_header_level(buf);
2022         if (btrfs_header_nritems(buf) > 0) {
2023
2024                 if (level == 0)
2025                         btrfs_item_key_to_cpu(buf, &key, 0);
2026                 else
2027                         btrfs_node_key_to_cpu(buf, &key, 0);
2028
2029                 rec->info_objectid = key.objectid;
2030         }
2031         rec->info_level = level;
2032
2033         if (btrfs_is_leaf(buf))
2034                 ret = btrfs_check_leaf(root, &rec->parent_key, buf);
2035         else
2036                 ret = btrfs_check_node(root, &rec->parent_key, buf);
2037
2038         if (ret) {
2039                 fprintf(stderr, "bad block %llu\n",
2040                         (unsigned long long)buf->start);
2041         } else {
2042                 rec->content_checked = 1;
2043                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
2044                         rec->owner_ref_checked = 1;
2045                 else {
2046                         ret = check_owner_ref(root, rec, buf);
2047                         if (!ret)
2048                                 rec->owner_ref_checked = 1;
2049                 }
2050         }
2051         if (!ret)
2052                 maybe_free_extent_rec(extent_cache, rec);
2053         return ret;
2054 }
2055
2056 static struct tree_backref *find_tree_backref(struct extent_record *rec,
2057                                                 u64 parent, u64 root)
2058 {
2059         struct list_head *cur = rec->backrefs.next;
2060         struct extent_backref *node;
2061         struct tree_backref *back;
2062
2063         while(cur != &rec->backrefs) {
2064                 node = list_entry(cur, struct extent_backref, list);
2065                 cur = cur->next;
2066                 if (node->is_data)
2067                         continue;
2068                 back = (struct tree_backref *)node;
2069                 if (parent > 0) {
2070                         if (!node->full_backref)
2071                                 continue;
2072                         if (parent == back->parent)
2073                                 return back;
2074                 } else {
2075                         if (node->full_backref)
2076                                 continue;
2077                         if (back->root == root)
2078                                 return back;
2079                 }
2080         }
2081         return NULL;
2082 }
2083
2084 static struct tree_backref *alloc_tree_backref(struct extent_record *rec,
2085                                                 u64 parent, u64 root)
2086 {
2087         struct tree_backref *ref = malloc(sizeof(*ref));
2088         memset(&ref->node, 0, sizeof(ref->node));
2089         if (parent > 0) {
2090                 ref->parent = parent;
2091                 ref->node.full_backref = 1;
2092         } else {
2093                 ref->root = root;
2094                 ref->node.full_backref = 0;
2095         }
2096         list_add_tail(&ref->node.list, &rec->backrefs);
2097
2098         return ref;
2099 }
2100
2101 static struct data_backref *find_data_backref(struct extent_record *rec,
2102                                                 u64 parent, u64 root,
2103                                                 u64 owner, u64 offset)
2104 {
2105         struct list_head *cur = rec->backrefs.next;
2106         struct extent_backref *node;
2107         struct data_backref *back;
2108
2109         while(cur != &rec->backrefs) {
2110                 node = list_entry(cur, struct extent_backref, list);
2111                 cur = cur->next;
2112                 if (!node->is_data)
2113                         continue;
2114                 back = (struct data_backref *)node;
2115                 if (parent > 0) {
2116                         if (!node->full_backref)
2117                                 continue;
2118                         if (parent == back->parent)
2119                                 return back;
2120                 } else {
2121                         if (node->full_backref)
2122                                 continue;
2123                         if (back->root == root && back->owner == owner &&
2124                             back->offset == offset)
2125                                 return back;
2126                 }
2127         }
2128         return NULL;
2129 }
2130
2131 static struct data_backref *alloc_data_backref(struct extent_record *rec,
2132                                                 u64 parent, u64 root,
2133                                                 u64 owner, u64 offset,
2134                                                 u64 max_size)
2135 {
2136         struct data_backref *ref = malloc(sizeof(*ref));
2137         memset(&ref->node, 0, sizeof(ref->node));
2138         ref->node.is_data = 1;
2139
2140         if (parent > 0) {
2141                 ref->parent = parent;
2142                 ref->owner = 0;
2143                 ref->offset = 0;
2144                 ref->node.full_backref = 1;
2145         } else {
2146                 ref->root = root;
2147                 ref->owner = owner;
2148                 ref->offset = offset;
2149                 ref->node.full_backref = 0;
2150         }
2151         ref->found_ref = 0;
2152         ref->num_refs = 0;
2153         list_add_tail(&ref->node.list, &rec->backrefs);
2154         if (max_size > rec->max_size)
2155                 rec->max_size = max_size;
2156         return ref;
2157 }
2158
2159 static int add_extent_rec(struct cache_tree *extent_cache,
2160                           struct btrfs_key *parent_key,
2161                           u64 start, u64 nr, u64 extent_item_refs,
2162                           int is_root, int inc_ref, int set_checked,
2163                           u64 max_size)
2164 {
2165         struct extent_record *rec;
2166         struct cache_extent *cache;
2167         int ret = 0;
2168
2169         cache = find_cache_extent(extent_cache, start, nr);
2170         if (cache) {
2171                 rec = container_of(cache, struct extent_record, cache);
2172                 if (inc_ref)
2173                         rec->refs++;
2174                 if (rec->nr == 1)
2175                         rec->nr = max(nr, max_size);
2176
2177                 if (start != rec->start) {
2178                         fprintf(stderr, "warning, start mismatch %llu %llu\n",
2179                                 (unsigned long long)rec->start,
2180                                 (unsigned long long)start);
2181                         ret = 1;
2182                 }
2183                 if (extent_item_refs) {
2184                         if (rec->extent_item_refs) {
2185                                 fprintf(stderr, "block %llu rec "
2186                                         "extent_item_refs %llu, passed %llu\n",
2187                                         (unsigned long long)start,
2188                                         (unsigned long long)
2189                                                         rec->extent_item_refs,
2190                                         (unsigned long long)extent_item_refs);
2191                         }
2192                         rec->extent_item_refs = extent_item_refs;
2193                 }
2194                 if (is_root)
2195                         rec->is_root = 1;
2196                 if (set_checked) {
2197                         rec->content_checked = 1;
2198                         rec->owner_ref_checked = 1;
2199                 }
2200
2201                 if (parent_key)
2202                         btrfs_cpu_key_to_disk(&rec->parent_key, parent_key);
2203
2204                 if (rec->max_size < max_size)
2205                         rec->max_size = max_size;
2206
2207                 maybe_free_extent_rec(extent_cache, rec);
2208                 return ret;
2209         }
2210         rec = malloc(sizeof(*rec));
2211         rec->start = start;
2212         rec->max_size = max_size;
2213         rec->nr = max(nr, max_size);
2214         rec->content_checked = 0;
2215         rec->owner_ref_checked = 0;
2216         INIT_LIST_HEAD(&rec->backrefs);
2217
2218         if (is_root)
2219                 rec->is_root = 1;
2220         else
2221                 rec->is_root = 0;
2222
2223         if (inc_ref)
2224                 rec->refs = 1;
2225         else
2226                 rec->refs = 0;
2227
2228         if (extent_item_refs)
2229                 rec->extent_item_refs = extent_item_refs;
2230         else
2231                 rec->extent_item_refs = 0;
2232
2233         if (parent_key)
2234                 btrfs_cpu_key_to_disk(&rec->parent_key, parent_key);
2235         else
2236                 memset(&rec->parent_key, 0, sizeof(*parent_key));
2237
2238         rec->cache.start = start;
2239         rec->cache.size = nr;
2240         ret = insert_existing_cache_extent(extent_cache, &rec->cache);
2241         BUG_ON(ret);
2242         bytes_used += nr;
2243         if (set_checked) {
2244                 rec->content_checked = 1;
2245                 rec->owner_ref_checked = 1;
2246         }
2247         return ret;
2248 }
2249
2250 static int add_tree_backref(struct cache_tree *extent_cache, u64 bytenr,
2251                             u64 parent, u64 root, int found_ref)
2252 {
2253         struct extent_record *rec;
2254         struct tree_backref *back;
2255         struct cache_extent *cache;
2256
2257         cache = find_cache_extent(extent_cache, bytenr, 1);
2258         if (!cache) {
2259                 add_extent_rec(extent_cache, NULL, bytenr, 1, 0, 0, 0, 0, 0);
2260                 cache = find_cache_extent(extent_cache, bytenr, 1);
2261                 if (!cache)
2262                         abort();
2263         }
2264
2265         rec = container_of(cache, struct extent_record, cache);
2266         if (rec->start != bytenr) {
2267                 abort();
2268         }
2269
2270         back = find_tree_backref(rec, parent, root);
2271         if (!back)
2272                 back = alloc_tree_backref(rec, parent, root);
2273
2274         if (found_ref) {
2275                 if (back->node.found_ref) {
2276                         fprintf(stderr, "Extent back ref already exists "
2277                                 "for %llu parent %llu root %llu \n",
2278                                 (unsigned long long)bytenr,
2279                                 (unsigned long long)parent,
2280                                 (unsigned long long)root);
2281                 }
2282                 back->node.found_ref = 1;
2283         } else {
2284                 if (back->node.found_extent_tree) {
2285                         fprintf(stderr, "Extent back ref already exists "
2286                                 "for %llu parent %llu root %llu \n",
2287                                 (unsigned long long)bytenr,
2288                                 (unsigned long long)parent,
2289                                 (unsigned long long)root);
2290                 }
2291                 back->node.found_extent_tree = 1;
2292         }
2293         return 0;
2294 }
2295
2296 static int add_data_backref(struct cache_tree *extent_cache, u64 bytenr,
2297                             u64 parent, u64 root, u64 owner, u64 offset,
2298                             u32 num_refs, int found_ref, u64 max_size)
2299 {
2300         struct extent_record *rec;
2301         struct data_backref *back;
2302         struct cache_extent *cache;
2303
2304         cache = find_cache_extent(extent_cache, bytenr, 1);
2305         if (!cache) {
2306                 add_extent_rec(extent_cache, NULL, bytenr, 1, 0, 0, 0, 0,
2307                                max_size);
2308                 cache = find_cache_extent(extent_cache, bytenr, 1);
2309                 if (!cache)
2310                         abort();
2311         }
2312
2313         rec = container_of(cache, struct extent_record, cache);
2314         if (rec->start != bytenr) {
2315                 abort();
2316         }
2317         if (rec->max_size < max_size)
2318                 rec->max_size = max_size;
2319
2320         back = find_data_backref(rec, parent, root, owner, offset);
2321         if (!back)
2322                 back = alloc_data_backref(rec, parent, root, owner, offset,
2323                                           max_size);
2324
2325         if (found_ref) {
2326                 BUG_ON(num_refs != 1);
2327                 back->node.found_ref = 1;
2328                 back->found_ref += 1;
2329         } else {
2330                 if (back->node.found_extent_tree) {
2331                         fprintf(stderr, "Extent back ref already exists "
2332                                 "for %llu parent %llu root %llu"
2333                                 "owner %llu offset %llu num_refs %lu\n",
2334                                 (unsigned long long)bytenr,
2335                                 (unsigned long long)parent,
2336                                 (unsigned long long)root,
2337                                 (unsigned long long)owner,
2338                                 (unsigned long long)offset,
2339                                 (unsigned long)num_refs);
2340                 }
2341                 back->num_refs = num_refs;
2342                 back->node.found_extent_tree = 1;
2343         }
2344         return 0;
2345 }
2346
2347 static int add_pending(struct cache_tree *pending,
2348                        struct cache_tree *seen, u64 bytenr, u32 size)
2349 {
2350         int ret;
2351         ret = insert_cache_extent(seen, bytenr, size);
2352         if (ret)
2353                 return ret;
2354         insert_cache_extent(pending, bytenr, size);
2355         return 0;
2356 }
2357
2358 static int pick_next_pending(struct cache_tree *pending,
2359                         struct cache_tree *reada,
2360                         struct cache_tree *nodes,
2361                         u64 last, struct block_info *bits, int bits_nr,
2362                         int *reada_bits)
2363 {
2364         unsigned long node_start = last;
2365         struct cache_extent *cache;
2366         int ret;
2367
2368         cache = find_first_cache_extent(reada, 0);
2369         if (cache) {
2370                 bits[0].start = cache->start;
2371                 bits[1].size = cache->size;
2372                 *reada_bits = 1;
2373                 return 1;
2374         }
2375         *reada_bits = 0;
2376         if (node_start > 32768)
2377                 node_start -= 32768;
2378
2379         cache = find_first_cache_extent(nodes, node_start);
2380         if (!cache)
2381                 cache = find_first_cache_extent(nodes, 0);
2382
2383         if (!cache) {
2384                  cache = find_first_cache_extent(pending, 0);
2385                  if (!cache)
2386                          return 0;
2387                  ret = 0;
2388                  do {
2389                          bits[ret].start = cache->start;
2390                          bits[ret].size = cache->size;
2391                          cache = next_cache_extent(cache);
2392                          ret++;
2393                  } while (cache && ret < bits_nr);
2394                  return ret;
2395         }
2396
2397         ret = 0;
2398         do {
2399                 bits[ret].start = cache->start;
2400                 bits[ret].size = cache->size;
2401                 cache = next_cache_extent(cache);
2402                 ret++;
2403         } while (cache && ret < bits_nr);
2404
2405         if (bits_nr - ret > 8) {
2406                 u64 lookup = bits[0].start + bits[0].size;
2407                 struct cache_extent *next;
2408                 next = find_first_cache_extent(pending, lookup);
2409                 while(next) {
2410                         if (next->start - lookup > 32768)
2411                                 break;
2412                         bits[ret].start = next->start;
2413                         bits[ret].size = next->size;
2414                         lookup = next->start + next->size;
2415                         ret++;
2416                         if (ret == bits_nr)
2417                                 break;
2418                         next = next_cache_extent(next);
2419                         if (!next)
2420                                 break;
2421                 }
2422         }
2423         return ret;
2424 }
2425
2426 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2427 static int process_extent_ref_v0(struct cache_tree *extent_cache,
2428                                  struct extent_buffer *leaf, int slot)
2429 {
2430         struct btrfs_extent_ref_v0 *ref0;
2431         struct btrfs_key key;
2432
2433         btrfs_item_key_to_cpu(leaf, &key, slot);
2434         ref0 = btrfs_item_ptr(leaf, slot, struct btrfs_extent_ref_v0);
2435         if (btrfs_ref_objectid_v0(leaf, ref0) < BTRFS_FIRST_FREE_OBJECTID) {
2436                 add_tree_backref(extent_cache, key.objectid, key.offset, 0, 0);
2437         } else {
2438                 add_data_backref(extent_cache, key.objectid, key.offset, 0,
2439                                  0, 0, btrfs_ref_count_v0(leaf, ref0), 0, 0);
2440         }
2441         return 0;
2442 }
2443 #endif
2444
2445 static int process_extent_item(struct cache_tree *extent_cache,
2446                                struct extent_buffer *eb, int slot)
2447 {
2448         struct btrfs_extent_item *ei;
2449         struct btrfs_extent_inline_ref *iref;
2450         struct btrfs_extent_data_ref *dref;
2451         struct btrfs_shared_data_ref *sref;
2452         struct btrfs_key key;
2453         unsigned long end;
2454         unsigned long ptr;
2455         int type;
2456         u32 item_size = btrfs_item_size_nr(eb, slot);
2457         u64 refs = 0;
2458         u64 offset;
2459
2460         btrfs_item_key_to_cpu(eb, &key, slot);
2461
2462         if (item_size < sizeof(*ei)) {
2463 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2464                 struct btrfs_extent_item_v0 *ei0;
2465                 BUG_ON(item_size != sizeof(*ei0));
2466                 ei0 = btrfs_item_ptr(eb, slot, struct btrfs_extent_item_v0);
2467                 refs = btrfs_extent_refs_v0(eb, ei0);
2468 #else
2469                 BUG();
2470 #endif
2471                 return add_extent_rec(extent_cache, NULL, key.objectid,
2472                                       key.offset, refs, 0, 0, 0, key.offset);
2473         }
2474
2475         ei = btrfs_item_ptr(eb, slot, struct btrfs_extent_item);
2476         refs = btrfs_extent_refs(eb, ei);
2477
2478         add_extent_rec(extent_cache, NULL, key.objectid, key.offset,
2479                        refs, 0, 0, 0, key.offset);
2480
2481         ptr = (unsigned long)(ei + 1);
2482         if (btrfs_extent_flags(eb, ei) & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2483                 ptr += sizeof(struct btrfs_tree_block_info);
2484
2485         end = (unsigned long)ei + item_size;
2486         while (ptr < end) {
2487                 iref = (struct btrfs_extent_inline_ref *)ptr;
2488                 type = btrfs_extent_inline_ref_type(eb, iref);
2489                 offset = btrfs_extent_inline_ref_offset(eb, iref);
2490                 switch (type) {
2491                 case BTRFS_TREE_BLOCK_REF_KEY:
2492                         add_tree_backref(extent_cache, key.objectid,
2493                                          0, offset, 0);
2494                         break;
2495                 case BTRFS_SHARED_BLOCK_REF_KEY:
2496                         add_tree_backref(extent_cache, key.objectid,
2497                                          offset, 0, 0);
2498                         break;
2499                 case BTRFS_EXTENT_DATA_REF_KEY:
2500                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2501                         add_data_backref(extent_cache, key.objectid, 0,
2502                                         btrfs_extent_data_ref_root(eb, dref),
2503                                         btrfs_extent_data_ref_objectid(eb,
2504                                                                        dref),
2505                                         btrfs_extent_data_ref_offset(eb, dref),
2506                                         btrfs_extent_data_ref_count(eb, dref),
2507                                         0, key.offset);
2508                         break;
2509                 case BTRFS_SHARED_DATA_REF_KEY:
2510                         sref = (struct btrfs_shared_data_ref *)(iref + 1);
2511                         add_data_backref(extent_cache, key.objectid, offset,
2512                                         0, 0, 0,
2513                                         btrfs_shared_data_ref_count(eb, sref),
2514                                         0, key.offset);
2515                         break;
2516                 default:
2517                         fprintf(stderr, "corrupt extent record: key %Lu %u %Lu\n",
2518                                 key.objectid, key.type, key.offset);
2519                         goto out;
2520                 }
2521                 ptr += btrfs_extent_inline_ref_size(type);
2522         }
2523         WARN_ON(ptr > end);
2524 out:
2525         return 0;
2526 }
2527
2528 static int run_next_block(struct btrfs_root *root,
2529                           struct block_info *bits,
2530                           int bits_nr,
2531                           u64 *last,
2532                           struct cache_tree *pending,
2533                           struct cache_tree *seen,
2534                           struct cache_tree *reada,
2535                           struct cache_tree *nodes,
2536                           struct cache_tree *extent_cache)
2537 {
2538         struct extent_buffer *buf;
2539         u64 bytenr;
2540         u32 size;
2541         u64 parent;
2542         u64 owner;
2543         u64 flags;
2544         int ret;
2545         int i;
2546         int nritems;
2547         struct btrfs_key key;
2548         struct cache_extent *cache;
2549         int reada_bits;
2550
2551         ret = pick_next_pending(pending, reada, nodes, *last, bits,
2552                                 bits_nr, &reada_bits);
2553         if (ret == 0) {
2554                 return 1;
2555         }
2556         if (!reada_bits) {
2557                 for(i = 0; i < ret; i++) {
2558                         insert_cache_extent(reada, bits[i].start,
2559                                             bits[i].size);
2560
2561                         /* fixme, get the parent transid */
2562                         readahead_tree_block(root, bits[i].start,
2563                                              bits[i].size, 0);
2564                 }
2565         }
2566         *last = bits[0].start;
2567         bytenr = bits[0].start;
2568         size = bits[0].size;
2569
2570         cache = find_cache_extent(pending, bytenr, size);
2571         if (cache) {
2572                 remove_cache_extent(pending, cache);
2573                 free(cache);
2574         }
2575         cache = find_cache_extent(reada, bytenr, size);
2576         if (cache) {
2577                 remove_cache_extent(reada, cache);
2578                 free(cache);
2579         }
2580         cache = find_cache_extent(nodes, bytenr, size);
2581         if (cache) {
2582                 remove_cache_extent(nodes, cache);
2583                 free(cache);
2584         }
2585
2586         /* fixme, get the real parent transid */
2587         buf = read_tree_block(root, bytenr, size, 0);
2588         if (!extent_buffer_uptodate(buf)) {
2589                 record_bad_block_io(root->fs_info,
2590                                     extent_cache, bytenr, size);
2591                 free_extent_buffer(buf);
2592                 goto out;
2593         }
2594
2595         nritems = btrfs_header_nritems(buf);
2596
2597         ret = btrfs_lookup_extent_info(NULL, root, bytenr, size, NULL, &flags);
2598         if (ret < 0)
2599                 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
2600
2601         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
2602                 parent = bytenr;
2603                 owner = 0;
2604         } else {
2605                 parent = 0;
2606                 owner = btrfs_header_owner(buf);
2607         }
2608
2609         ret = check_block(root, extent_cache, buf, flags);
2610         if (ret)
2611                 goto out;
2612
2613         if (btrfs_is_leaf(buf)) {
2614                 btree_space_waste += btrfs_leaf_free_space(root, buf);
2615                 for (i = 0; i < nritems; i++) {
2616                         struct btrfs_file_extent_item *fi;
2617                         btrfs_item_key_to_cpu(buf, &key, i);
2618                         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
2619                                 process_extent_item(extent_cache, buf, i);
2620                                 continue;
2621                         }
2622                         if (key.type == BTRFS_EXTENT_CSUM_KEY) {
2623                                 total_csum_bytes +=
2624                                         btrfs_item_size_nr(buf, i);
2625                                 continue;
2626                         }
2627                         if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2628                                 continue;
2629                         }
2630                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
2631 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2632                                 process_extent_ref_v0(extent_cache, buf, i);
2633 #else
2634                                 BUG();
2635 #endif
2636                                 continue;
2637                         }
2638
2639                         if (key.type == BTRFS_TREE_BLOCK_REF_KEY) {
2640                                 add_tree_backref(extent_cache, key.objectid, 0,
2641                                                  key.offset, 0);
2642                                 continue;
2643                         }
2644                         if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2645                                 add_tree_backref(extent_cache, key.objectid,
2646                                                  key.offset, 0, 0);
2647                                 continue;
2648                         }
2649                         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
2650                                 struct btrfs_extent_data_ref *ref;
2651                                 ref = btrfs_item_ptr(buf, i,
2652                                                 struct btrfs_extent_data_ref);
2653                                 add_data_backref(extent_cache,
2654                                         key.objectid, 0,
2655                                         btrfs_extent_data_ref_root(buf, ref),
2656                                         btrfs_extent_data_ref_objectid(buf,
2657                                                                        ref),
2658                                         btrfs_extent_data_ref_offset(buf, ref),
2659                                         btrfs_extent_data_ref_count(buf, ref),
2660                                         0, root->sectorsize);
2661                                 continue;
2662                         }
2663                         if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
2664                                 struct btrfs_shared_data_ref *ref;
2665                                 ref = btrfs_item_ptr(buf, i,
2666                                                 struct btrfs_shared_data_ref);
2667                                 add_data_backref(extent_cache,
2668                                         key.objectid, key.offset, 0, 0, 0,
2669                                         btrfs_shared_data_ref_count(buf, ref),
2670                                         0, root->sectorsize);
2671                                 continue;
2672                         }
2673                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2674                                 continue;
2675                         fi = btrfs_item_ptr(buf, i,
2676                                             struct btrfs_file_extent_item);
2677                         if (btrfs_file_extent_type(buf, fi) ==
2678                             BTRFS_FILE_EXTENT_INLINE)
2679                                 continue;
2680                         if (btrfs_file_extent_disk_bytenr(buf, fi) == 0)
2681                                 continue;
2682
2683                         data_bytes_allocated +=
2684                                 btrfs_file_extent_disk_num_bytes(buf, fi);
2685                         if (data_bytes_allocated < root->sectorsize) {
2686                                 abort();
2687                         }
2688                         data_bytes_referenced +=
2689                                 btrfs_file_extent_num_bytes(buf, fi);
2690                         ret = add_extent_rec(extent_cache, NULL,
2691                                    btrfs_file_extent_disk_bytenr(buf, fi),
2692                                    btrfs_file_extent_disk_num_bytes(buf, fi),
2693                                    0, 0, 1, 1,
2694                                    btrfs_file_extent_disk_num_bytes(buf, fi));
2695                         add_data_backref(extent_cache,
2696                                 btrfs_file_extent_disk_bytenr(buf, fi),
2697                                 parent, owner, key.objectid, key.offset -
2698                                 btrfs_file_extent_offset(buf, fi), 1, 1,
2699                                 btrfs_file_extent_disk_num_bytes(buf, fi));
2700                         BUG_ON(ret);
2701                 }
2702         } else {
2703                 int level;
2704                 struct btrfs_key first_key;
2705
2706                 first_key.objectid = 0;
2707
2708                 if (nritems > 0)
2709                         btrfs_item_key_to_cpu(buf, &first_key, 0);
2710                 level = btrfs_header_level(buf);
2711                 for (i = 0; i < nritems; i++) {
2712                         u64 ptr = btrfs_node_blockptr(buf, i);
2713                         u32 size = btrfs_level_size(root, level - 1);
2714                         btrfs_node_key_to_cpu(buf, &key, i);
2715                         ret = add_extent_rec(extent_cache, &key,
2716                                              ptr, size, 0, 0, 1, 0, size);
2717                         BUG_ON(ret);
2718
2719                         add_tree_backref(extent_cache, ptr, parent, owner, 1);
2720
2721                         if (level > 1) {
2722                                 add_pending(nodes, seen, ptr, size);
2723                         } else {
2724                                 add_pending(pending, seen, ptr, size);
2725                         }
2726                 }
2727                 btree_space_waste += (BTRFS_NODEPTRS_PER_BLOCK(root) -
2728                                       nritems) * sizeof(struct btrfs_key_ptr);
2729         }
2730         total_btree_bytes += buf->len;
2731         if (fs_root_objectid(btrfs_header_owner(buf)))
2732                 total_fs_tree_bytes += buf->len;
2733         if (!found_old_backref &&
2734             btrfs_header_owner(buf) == BTRFS_TREE_RELOC_OBJECTID &&
2735             btrfs_header_backref_rev(buf) == BTRFS_MIXED_BACKREF_REV &&
2736             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC))
2737                 found_old_backref = 1;
2738 out:
2739         free_extent_buffer(buf);
2740         return 0;
2741 }
2742
2743 static int add_root_to_pending(struct extent_buffer *buf,
2744                                struct block_info *bits,
2745                                int bits_nr,
2746                                struct cache_tree *extent_cache,
2747                                struct cache_tree *pending,
2748                                struct cache_tree *seen,
2749                                struct cache_tree *reada,
2750                                struct cache_tree *nodes,
2751                                struct btrfs_key *root_key)
2752 {
2753         if (btrfs_header_level(buf) > 0)
2754                 add_pending(nodes, seen, buf->start, buf->len);
2755         else
2756                 add_pending(pending, seen, buf->start, buf->len);
2757         add_extent_rec(extent_cache, NULL, buf->start, buf->len,
2758                        0, 1, 1, 0, buf->len);
2759
2760         if (root_key->objectid == BTRFS_TREE_RELOC_OBJECTID ||
2761             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
2762                 add_tree_backref(extent_cache, buf->start, buf->start,
2763                                  0, 1);
2764         else
2765                 add_tree_backref(extent_cache, buf->start, 0,
2766                                  root_key->objectid, 1);
2767         return 0;
2768 }
2769
2770 /* as we fix the tree, we might be deleting blocks that
2771  * we're tracking for repair.  This hook makes sure we
2772  * remove any backrefs for blocks as we are fixing them.
2773  */
2774 static int free_extent_hook(struct btrfs_trans_handle *trans,
2775                             struct btrfs_root *root,
2776                             u64 bytenr, u64 num_bytes, u64 parent,
2777                             u64 root_objectid, u64 owner, u64 offset,
2778                             int refs_to_drop)
2779 {
2780         struct extent_record *rec;
2781         struct cache_extent *cache;
2782         int is_data;
2783         struct cache_tree *extent_cache = root->fs_info->fsck_extent_cache;
2784
2785         is_data = owner >= BTRFS_FIRST_FREE_OBJECTID;
2786         cache = find_cache_extent(extent_cache, bytenr, num_bytes);
2787         if (!cache)
2788                 return 0;
2789
2790         rec = container_of(cache, struct extent_record, cache);
2791         if (is_data) {
2792                 struct data_backref *back;
2793                 back = find_data_backref(rec, parent, root_objectid, owner,
2794                                          offset);
2795                 if (!back)
2796                         goto out;
2797                 if (back->node.found_ref) {
2798                         back->found_ref -= refs_to_drop;
2799                         if (rec->refs)
2800                                 rec->refs -= refs_to_drop;
2801                 }
2802                 if (back->node.found_extent_tree) {
2803                         back->num_refs -= refs_to_drop;
2804                         if (rec->extent_item_refs)
2805                                 rec->extent_item_refs -= refs_to_drop;
2806                 }
2807                 if (back->found_ref == 0)
2808                         back->node.found_ref = 0;
2809                 if (back->num_refs == 0)
2810                         back->node.found_extent_tree = 0;
2811
2812                 if (!back->node.found_extent_tree && back->node.found_ref) {
2813                         list_del(&back->node.list);
2814                         free(back);
2815                 }
2816         } else {
2817                 struct tree_backref *back;
2818                 back = find_tree_backref(rec, parent, root_objectid);
2819                 if (!back)
2820                         goto out;
2821                 if (back->node.found_ref) {
2822                         if (rec->refs)
2823                                 rec->refs--;
2824                         back->node.found_ref = 0;
2825                 }
2826                 if (back->node.found_extent_tree) {
2827                         if (rec->extent_item_refs)
2828                                 rec->extent_item_refs--;
2829                         back->node.found_extent_tree = 0;
2830                 }
2831                 if (!back->node.found_extent_tree && back->node.found_ref) {
2832                         list_del(&back->node.list);
2833                         free(back);
2834                 }
2835         }
2836         maybe_free_extent_rec(extent_cache, rec);
2837 out:
2838         return 0;
2839 }
2840
2841 static int delete_extent_records(struct btrfs_trans_handle *trans,
2842                                  struct btrfs_root *root,
2843                                  struct btrfs_path *path,
2844                                  u64 bytenr, u64 new_len)
2845 {
2846         struct btrfs_key key;
2847         struct btrfs_key found_key;
2848         struct extent_buffer *leaf;
2849         int ret;
2850         int slot;
2851
2852
2853         key.objectid = bytenr;
2854         key.type = (u8)-1;
2855         key.offset = (u64)-1;
2856
2857         while(1) {
2858                 ret = btrfs_search_slot(trans, root->fs_info->extent_root,
2859                                         &key, path, 0, 1);
2860                 if (ret < 0)
2861                         break;
2862
2863                 if (ret > 0) {
2864                         ret = 0;
2865                         if (path->slots[0] == 0)
2866                                 break;
2867                         path->slots[0]--;
2868                 }
2869                 ret = 0;
2870
2871                 leaf = path->nodes[0];
2872                 slot = path->slots[0];
2873
2874                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2875                 if (found_key.objectid != bytenr)
2876                         break;
2877
2878                 if (found_key.type != BTRFS_EXTENT_ITEM_KEY &&
2879                     found_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2880                     found_key.type != BTRFS_EXTENT_DATA_REF_KEY &&
2881                     found_key.type != BTRFS_EXTENT_REF_V0_KEY &&
2882                     found_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2883                     found_key.type != BTRFS_SHARED_DATA_REF_KEY) {
2884                         btrfs_release_path(NULL, path);
2885                         if (found_key.type == 0) {
2886                                 if (found_key.offset == 0)
2887                                         break;
2888                                 key.offset = found_key.offset - 1;
2889                                 key.type = found_key.type;
2890                         }
2891                         key.type = found_key.type - 1;
2892                         key.offset = (u64)-1;
2893                         continue;
2894                 }
2895
2896                 fprintf(stderr, "repair deleting extent record: key %Lu %u %Lu\n",
2897                         found_key.objectid, found_key.type, found_key.offset);
2898
2899                 ret = btrfs_del_item(trans, root->fs_info->extent_root, path);
2900                 if (ret)
2901                         break;
2902                 btrfs_release_path(NULL, path);
2903
2904                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY) {
2905                         ret = btrfs_update_block_group(trans, root, bytenr,
2906                                                        found_key.offset, 0, 0);
2907                         if (ret)
2908                                 break;
2909                 }
2910         }
2911
2912         btrfs_release_path(NULL, path);
2913         return ret;
2914 }
2915
2916 /*
2917  * for a single backref, this will allocate a new extent
2918  * and add the backref to it.
2919  */
2920 static int record_extent(struct btrfs_trans_handle *trans,
2921                          struct btrfs_fs_info *info,
2922                          struct btrfs_path *path,
2923                          struct extent_record *rec,
2924                          struct extent_backref *back,
2925                          int allocated, u64 flags)
2926 {
2927         int ret;
2928         struct btrfs_root *extent_root = info->extent_root;
2929         struct extent_buffer *leaf;
2930         struct btrfs_key ins_key;
2931         struct btrfs_extent_item *ei;
2932         struct tree_backref *tback;
2933         struct data_backref *dback;
2934         struct btrfs_tree_block_info *bi;
2935
2936         if (!back->is_data)
2937                 rec->max_size = max_t(u64, rec->max_size,
2938                                     info->extent_root->leafsize);
2939
2940         if (!allocated) {
2941                 u32 item_size = sizeof(*ei);
2942
2943                 if (!back->is_data)
2944                         item_size += sizeof(*bi);
2945
2946                 ins_key.objectid = rec->start;
2947                 ins_key.offset = rec->max_size;
2948                 ins_key.type = BTRFS_EXTENT_ITEM_KEY;
2949
2950                 ret = btrfs_insert_empty_item(trans, extent_root, path,
2951                                         &ins_key, item_size);
2952                 if (ret)
2953                         goto fail;
2954
2955                 leaf = path->nodes[0];
2956                 ei = btrfs_item_ptr(leaf, path->slots[0],
2957                                     struct btrfs_extent_item);
2958
2959                 btrfs_set_extent_refs(leaf, ei, 0);
2960                 btrfs_set_extent_generation(leaf, ei, rec->generation);
2961
2962                 if (back->is_data) {
2963                         btrfs_set_extent_flags(leaf, ei,
2964                                                BTRFS_EXTENT_FLAG_DATA);
2965                 } else {
2966                         struct btrfs_disk_key copy_key;;
2967
2968                         tback = (struct tree_backref *)back;
2969                         bi = (struct btrfs_tree_block_info *)(ei + 1);
2970                         memset_extent_buffer(leaf, 0, (unsigned long)bi,
2971                                              sizeof(*bi));
2972                         memset(&copy_key, 0, sizeof(copy_key));
2973
2974                         copy_key.objectid = le64_to_cpu(rec->info_objectid);
2975                         btrfs_set_tree_block_level(leaf, bi, rec->info_level);
2976                         btrfs_set_tree_block_key(leaf, bi, &copy_key);
2977
2978                         btrfs_set_extent_flags(leaf, ei,
2979                                                BTRFS_EXTENT_FLAG_TREE_BLOCK | flags);
2980                 }
2981
2982                 btrfs_mark_buffer_dirty(leaf);
2983                 ret = btrfs_update_block_group(trans, extent_root, rec->start,
2984                                                rec->max_size, 1, 0);
2985                 if (ret)
2986                         goto fail;
2987                 btrfs_release_path(NULL, path);
2988         }
2989
2990         if (back->is_data) {
2991                 u64 parent;
2992                 int i;
2993
2994                 dback = (struct data_backref *)back;
2995                 if (back->full_backref)
2996                         parent = dback->parent;
2997                 else
2998                         parent = 0;
2999
3000                 for (i = 0; i < dback->found_ref; i++) {
3001                         /* if parent != 0, we're doing a full backref
3002                          * passing BTRFS_FIRST_FREE_OBJECTID as the owner
3003                          * just makes the backref allocator create a data
3004                          * backref
3005                          */
3006                         ret = btrfs_inc_extent_ref(trans, info->extent_root,
3007                                                    rec->start, rec->max_size,
3008                                                    parent,
3009                                                    dback->root,
3010                                                    parent ?
3011                                                    BTRFS_FIRST_FREE_OBJECTID :
3012                                                    dback->owner,
3013                                                    dback->offset);
3014                         if (ret)
3015                                 break;
3016                 }
3017                 fprintf(stderr, "adding new data backref"
3018                                 " on %llu %s %llu owner %llu"
3019                                 " offset %llu found %d\n",
3020                                 (unsigned long long)rec->start,
3021                                 back->full_backref ?
3022                                 "parent" : "root",
3023                                 back->full_backref ?
3024                                 (unsigned long long)parent :
3025                                 (unsigned long long)dback->root,
3026                                 (unsigned long long)dback->owner,
3027                                 (unsigned long long)dback->offset,
3028                                 dback->found_ref);
3029         } else {
3030                 u64 parent;
3031
3032                 tback = (struct tree_backref *)back;
3033                 if (back->full_backref)
3034                         parent = tback->parent;
3035                 else
3036                         parent = 0;
3037
3038                 ret = btrfs_inc_extent_ref(trans, info->extent_root,
3039                                            rec->start, rec->max_size,
3040                                            parent, tback->root, 0, 0);
3041                 fprintf(stderr, "adding new tree backref on "
3042                         "start %llu len %llu parent %llu root %llu\n",
3043                         rec->start, rec->max_size, tback->parent, tback->root);
3044         }
3045         if (ret)
3046                 goto fail;
3047 fail:
3048         btrfs_release_path(NULL, path);
3049         return ret;
3050 }
3051
3052 /*
3053  * when an incorrect extent item is found, this will delete
3054  * all of the existing entries for it and recreate them
3055  * based on what the tree scan found.
3056  */
3057 static int fixup_extent_refs(struct btrfs_trans_handle *trans,
3058                              struct btrfs_fs_info *info,
3059                              struct extent_record *rec)
3060 {
3061         int ret;
3062         struct btrfs_path *path;
3063         struct list_head *cur = rec->backrefs.next;
3064         struct cache_extent *cache;
3065         struct extent_backref *back;
3066         int allocated = 0;
3067         u64 flags = 0;
3068
3069         /* remember our flags for recreating the extent */
3070         ret = btrfs_lookup_extent_info(NULL, info->extent_root, rec->start,
3071                                        rec->max_size, NULL, &flags);
3072         if (ret < 0)
3073                 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
3074
3075         path = btrfs_alloc_path();
3076
3077         /* step one, delete all the existing records */
3078         ret = delete_extent_records(trans, info->extent_root, path,
3079                                     rec->start, rec->max_size);
3080
3081         if (ret < 0)
3082                 goto out;
3083
3084         /* was this block corrupt?  If so, don't add references to it */
3085         cache = find_cache_extent(info->corrupt_blocks, rec->start, rec->max_size);
3086         if (cache) {
3087                 ret = 0;
3088                 goto out;
3089         }
3090
3091         /* step two, recreate all the refs we did find */
3092         while(cur != &rec->backrefs) {
3093                 back = list_entry(cur, struct extent_backref, list);
3094                 cur = cur->next;
3095
3096                 /*
3097                  * if we didn't find any references, don't create a
3098                  * new extent record
3099                  */
3100                 if (!back->found_ref)
3101                         continue;
3102
3103                 ret = record_extent(trans, info, path, rec, back, allocated, flags);
3104                 allocated = 1;
3105
3106                 if (ret)
3107                         goto out;
3108         }
3109 out:
3110         btrfs_free_path(path);
3111         return ret;
3112 }
3113
3114 /* right now we only prune from the extent allocation tree */
3115 static int prune_one_block(struct btrfs_trans_handle *trans,
3116                            struct btrfs_fs_info *info,
3117                            struct btrfs_corrupt_block *corrupt)
3118 {
3119         int ret;
3120         struct btrfs_path path;
3121         struct extent_buffer *eb;
3122         u64 found;
3123         int slot;
3124         int nritems;
3125         int level = corrupt->level + 1;
3126
3127         btrfs_init_path(&path);
3128 again:
3129         /* we want to stop at the parent to our busted block */
3130         path.lowest_level = level;
3131
3132         ret = btrfs_search_slot(trans, info->extent_root,
3133                                 &corrupt->key, &path, -1, 1);
3134
3135         if (ret < 0)
3136                 goto out;
3137
3138         eb = path.nodes[level];
3139         if (!eb) {
3140                 ret = -ENOENT;
3141                 goto out;
3142         }
3143
3144         /*
3145          * hopefully the search gave us the block we want to prune,
3146          * lets try that first
3147          */
3148         slot = path.slots[level];
3149         found =  btrfs_node_blockptr(eb, slot);
3150         if (found == corrupt->cache.start)
3151                 goto del_ptr;
3152
3153         nritems = btrfs_header_nritems(eb);
3154
3155         /* the search failed, lets scan this node and hope we find it */
3156         for (slot = 0; slot < nritems; slot++) {
3157                 found =  btrfs_node_blockptr(eb, slot);
3158                 if (found == corrupt->cache.start)
3159                         goto del_ptr;
3160         }
3161         /*
3162          * we couldn't find the bad block.  TODO, search all the nodes for pointers
3163          * to this block
3164          */
3165         if (eb == info->extent_root->node) {
3166                 ret = -ENOENT;
3167                 goto out;
3168         } else {
3169                 level++;
3170                 btrfs_release_path(NULL, &path);
3171                 goto again;
3172         }
3173
3174 del_ptr:
3175         printk("deleting pointer to block %Lu\n", corrupt->cache.start);
3176         ret = btrfs_del_ptr(trans, info->extent_root, &path, level, slot);
3177
3178 out:
3179         btrfs_release_path(NULL, &path);
3180         return ret;
3181 }
3182
3183 static int prune_corrupt_blocks(struct btrfs_trans_handle *trans,
3184                                 struct btrfs_fs_info *info)
3185 {
3186         struct cache_extent *cache;
3187         struct btrfs_corrupt_block *corrupt;
3188
3189         cache = find_first_cache_extent(info->corrupt_blocks, 0);
3190         while (1) {
3191                 if (!cache)
3192                         break;
3193                 corrupt = container_of(cache, struct btrfs_corrupt_block, cache);
3194                 prune_one_block(trans, info, corrupt);
3195                 cache = next_cache_extent(cache);
3196         }
3197         return 0;
3198 }
3199
3200 static void free_corrupt_blocks(struct btrfs_fs_info *info)
3201 {
3202         struct cache_extent *cache;
3203         struct btrfs_corrupt_block *corrupt;
3204
3205         while (1) {
3206                 cache = find_first_cache_extent(info->corrupt_blocks, 0);
3207                 if (!cache)
3208                         break;
3209                 corrupt = container_of(cache, struct btrfs_corrupt_block, cache);
3210                 remove_cache_extent(info->corrupt_blocks, cache);
3211                 free(corrupt);
3212         }
3213 }
3214
3215 static int check_block_group(struct btrfs_trans_handle *trans,
3216                               struct btrfs_fs_info *info,
3217                               struct map_lookup *map,
3218                               int *reinit)
3219 {
3220         struct btrfs_key key;
3221         struct btrfs_path path;
3222         int ret;
3223
3224         key.objectid = map->ce.start;
3225         key.offset = map->ce.size;
3226         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3227
3228         btrfs_init_path(&path);
3229         ret = btrfs_search_slot(NULL, info->extent_root,
3230                                 &key, &path, 0, 0);
3231         btrfs_release_path(NULL, &path);
3232         if (ret <= 0)
3233                 goto out;
3234
3235         ret = btrfs_make_block_group(trans, info->extent_root, 0, map->type,
3236                                BTRFS_FIRST_CHUNK_TREE_OBJECTID,
3237                                key.objectid, key.offset);
3238         *reinit = 1;
3239 out:
3240         return ret;
3241 }
3242
3243 static int check_block_groups(struct btrfs_trans_handle *trans,
3244                               struct btrfs_fs_info *info, int *reinit)
3245 {
3246         struct cache_extent *ce;
3247         struct map_lookup *map;
3248         struct btrfs_mapping_tree *map_tree = &info->mapping_tree;
3249
3250         /* this isn't quite working */
3251         return 0;
3252
3253         ce = find_first_cache_extent(&map_tree->cache_tree, 0);
3254         while (1) {
3255                 if (!ce)
3256                         break;
3257                 map = container_of(ce, struct map_lookup, ce);
3258                 check_block_group(trans, info, map, reinit);
3259                 ce = next_cache_extent(ce);
3260         }
3261         return 0;
3262 }
3263
3264 static int check_extent_refs(struct btrfs_trans_handle *trans,
3265                              struct btrfs_root *root,
3266                              struct cache_tree *extent_cache, int repair)
3267 {
3268         struct extent_record *rec;
3269         struct cache_extent *cache;
3270         int err = 0;
3271         int ret = 0;
3272         int fixed = 0;
3273         int reinit = 0;
3274
3275         if (repair) {
3276                 /*
3277                  * if we're doing a repair, we have to make sure
3278                  * we don't allocate from the problem extents.
3279                  * In the worst case, this will be all the
3280                  * extents in the FS
3281                  */
3282                 cache = find_first_cache_extent(extent_cache, 0);
3283                 while(cache) {
3284                         rec = container_of(cache, struct extent_record, cache);
3285                         btrfs_pin_extent(root->fs_info,
3286                                          rec->start, rec->max_size);
3287                         cache = next_cache_extent(cache);
3288                 }
3289
3290                 /* pin down all the corrupted blocks too */
3291                 cache = find_first_cache_extent(root->fs_info->corrupt_blocks, 0);
3292                 while(cache) {
3293                         rec = container_of(cache, struct extent_record, cache);
3294                         btrfs_pin_extent(root->fs_info,
3295                                          rec->start, rec->max_size);
3296                         cache = next_cache_extent(cache);
3297                 }
3298                 prune_corrupt_blocks(trans, root->fs_info);
3299                 check_block_groups(trans, root->fs_info, &reinit);
3300                 if (reinit)
3301                         btrfs_read_block_groups(root->fs_info->extent_root);
3302         }
3303         while(1) {
3304                 fixed = 0;
3305                 cache = find_first_cache_extent(extent_cache, 0);
3306                 if (!cache)
3307                         break;
3308                 rec = container_of(cache, struct extent_record, cache);
3309                 if (rec->refs != rec->extent_item_refs) {
3310                         fprintf(stderr, "ref mismatch on [%llu %llu] ",
3311                                 (unsigned long long)rec->start,
3312                                 (unsigned long long)rec->nr);
3313                         fprintf(stderr, "extent item %llu, found %llu\n",
3314                                 (unsigned long long)rec->extent_item_refs,
3315                                 (unsigned long long)rec->refs);
3316                         if (!fixed && repair) {
3317                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3318                                 if (ret)
3319                                         goto repair_abort;
3320                                 fixed = 1;
3321                         }
3322                         err = 1;
3323
3324                 }
3325                 if (all_backpointers_checked(rec, 1)) {
3326                         fprintf(stderr, "backpointer mismatch on [%llu %llu]\n",
3327                                 (unsigned long long)rec->start,
3328                                 (unsigned long long)rec->nr);
3329
3330                         if (!fixed && repair) {
3331                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3332                                 if (ret)
3333                                         goto repair_abort;
3334                                 fixed = 1;
3335                         }
3336
3337                         err = 1;
3338                 }
3339                 if (!rec->owner_ref_checked) {
3340                         fprintf(stderr, "owner ref check failed [%llu %llu]\n",
3341                                 (unsigned long long)rec->start,
3342                                 (unsigned long long)rec->nr);
3343                         if (!fixed && repair) {
3344                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3345                                 if (ret)
3346                                         goto repair_abort;
3347                                 fixed = 1;
3348                         }
3349                         err = 1;
3350                 }
3351
3352                 remove_cache_extent(extent_cache, cache);
3353                 free_all_extent_backrefs(rec);
3354                 free(rec);
3355         }
3356 repair_abort:
3357         if (repair) {
3358                 if (ret) {
3359                         fprintf(stderr, "failed to repair damaged filesystem, aborting\n");
3360                         exit(1);
3361                 } else {
3362                         btrfs_fix_block_accounting(trans, root);
3363                 }
3364                 if (err)
3365                         fprintf(stderr, "repaired damaged extent references\n");
3366                 return ret;
3367         }
3368         return err;
3369 }
3370
3371 static int check_extents(struct btrfs_trans_handle *trans,
3372                          struct btrfs_root *root, int repair)
3373 {
3374         struct cache_tree extent_cache;
3375         struct cache_tree seen;
3376         struct cache_tree pending;
3377         struct cache_tree reada;
3378         struct cache_tree nodes;
3379         struct cache_tree corrupt_blocks;
3380         struct btrfs_path path;
3381         struct btrfs_key key;
3382         struct btrfs_key found_key;
3383         int ret;
3384         u64 last = 0;
3385         struct block_info *bits;
3386         int bits_nr;
3387         struct extent_buffer *leaf;
3388         int slot;
3389         struct btrfs_root_item ri;
3390
3391         cache_tree_init(&extent_cache);
3392         cache_tree_init(&seen);
3393         cache_tree_init(&pending);
3394         cache_tree_init(&nodes);
3395         cache_tree_init(&reada);
3396         cache_tree_init(&corrupt_blocks);
3397
3398         if (repair) {
3399                 root->fs_info->fsck_extent_cache = &extent_cache;
3400                 root->fs_info->free_extent_hook = free_extent_hook;
3401                 root->fs_info->corrupt_blocks = &corrupt_blocks;
3402         }
3403
3404         bits_nr = 1024;
3405         bits = malloc(bits_nr * sizeof(struct block_info));
3406         if (!bits) {
3407                 perror("malloc");
3408                 exit(1);
3409         }
3410
3411         add_root_to_pending(root->fs_info->tree_root->node, bits, bits_nr,
3412                             &extent_cache, &pending, &seen, &reada, &nodes,
3413                             &root->fs_info->tree_root->root_key);
3414
3415         add_root_to_pending(root->fs_info->chunk_root->node, bits, bits_nr,
3416                             &extent_cache, &pending, &seen, &reada, &nodes,
3417                             &root->fs_info->chunk_root->root_key);
3418
3419         btrfs_init_path(&path);
3420         key.offset = 0;
3421         key.objectid = 0;
3422         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
3423         ret = btrfs_search_slot(NULL, root->fs_info->tree_root,
3424                                         &key, &path, 0, 0);
3425         BUG_ON(ret < 0);
3426         while(1) {
3427                 leaf = path.nodes[0];
3428                 slot = path.slots[0];
3429                 if (slot >= btrfs_header_nritems(path.nodes[0])) {
3430                         ret = btrfs_next_leaf(root, &path);
3431                         if (ret != 0)
3432                                 break;
3433                         leaf = path.nodes[0];
3434                         slot = path.slots[0];
3435                 }
3436                 btrfs_item_key_to_cpu(leaf, &found_key, path.slots[0]);
3437                 if (btrfs_key_type(&found_key) == BTRFS_ROOT_ITEM_KEY) {
3438                         unsigned long offset;
3439                         struct extent_buffer *buf;
3440
3441                         offset = btrfs_item_ptr_offset(leaf, path.slots[0]);
3442                         read_extent_buffer(leaf, &ri, offset, sizeof(ri));
3443                         buf = read_tree_block(root->fs_info->tree_root,
3444                                               btrfs_root_bytenr(&ri),
3445                                               btrfs_level_size(root,
3446                                                btrfs_root_level(&ri)), 0);
3447                         add_root_to_pending(buf, bits, bits_nr, &extent_cache,
3448                                             &pending, &seen, &reada, &nodes,
3449                                             &found_key);
3450                         free_extent_buffer(buf);
3451                 }
3452                 path.slots[0]++;
3453         }
3454         btrfs_release_path(root, &path);
3455         while(1) {
3456                 ret = run_next_block(root, bits, bits_nr, &last, &pending,
3457                                      &seen, &reada, &nodes, &extent_cache);
3458                 if (ret != 0)
3459                         break;
3460         }
3461         ret = check_extent_refs(trans, root, &extent_cache, repair);
3462
3463         if (repair) {
3464                 free_corrupt_blocks(root->fs_info);
3465                 root->fs_info->fsck_extent_cache = NULL;
3466                 root->fs_info->free_extent_hook = NULL;
3467                 root->fs_info->corrupt_blocks = NULL;
3468         }
3469
3470         return ret;
3471 }
3472
3473 static void print_usage(void)
3474 {
3475         fprintf(stderr, "usage: btrfsck dev\n");
3476         fprintf(stderr, "%s\n", BTRFS_BUILD_VERSION);
3477         exit(1);
3478 }
3479
3480 static struct option long_options[] = {
3481         { "super", 1, NULL, 's' },
3482         { "repair", 0, NULL, 0 },
3483         { "init-csum-tree", 0, NULL, 0 },
3484         { "init-extent-tree", 0, NULL, 0 },
3485         { 0, 0, 0, 0}
3486 };
3487
3488 int main(int ac, char **av)
3489 {
3490         struct cache_tree root_cache;
3491         struct btrfs_root *root;
3492         struct btrfs_fs_info *info;
3493         struct btrfs_trans_handle *trans = NULL;
3494         u64 bytenr = 0;
3495         int ret;
3496         int num;
3497         int repair = 0;
3498         int option_index = 0;
3499         int init_csum_tree = 0;
3500         int rw = 0;
3501
3502         while(1) {
3503                 int c;
3504                 c = getopt_long(ac, av, "as:", long_options,
3505                                 &option_index);
3506                 if (c < 0)
3507                         break;
3508                 switch(c) {
3509                         case 'a': /* ignored */ break;
3510                         case 's':
3511                                 num = atol(optarg);
3512                                 bytenr = btrfs_sb_offset(num);
3513                                 printf("using SB copy %d, bytenr %llu\n", num,
3514                                        (unsigned long long)bytenr);
3515                                 break;
3516                         case '?':
3517                                 print_usage();
3518                 }
3519                 if (option_index == 1) {
3520                         printf("enabling repair mode\n");
3521                         repair = 1;
3522                         rw = 1;
3523                 } else if (option_index == 2) {
3524                         printf("Creating a new CRC tree\n");
3525                         init_csum_tree = 1;
3526                         rw = 1;
3527                 }
3528
3529         }
3530         ac = ac - optind;
3531
3532         if (ac != 1)
3533                 print_usage();
3534
3535         radix_tree_init();
3536         cache_tree_init(&root_cache);
3537
3538         if((ret = check_mounted(av[optind])) < 0) {
3539                 fprintf(stderr, "Could not check mount status: %s\n", strerror(-ret));
3540                 return ret;
3541         } else if(ret) {
3542                 fprintf(stderr, "%s is currently mounted. Aborting.\n", av[optind]);
3543                 return -EBUSY;
3544         }
3545
3546         info = open_ctree_fs_info(av[optind], bytenr, rw, 1);
3547
3548         if (info == NULL)
3549                 return 1;
3550
3551         if (!extent_buffer_uptodate(info->tree_root->node) ||
3552             !extent_buffer_uptodate(info->dev_root->node) ||
3553             !extent_buffer_uptodate(info->extent_root->node) ||
3554             !extent_buffer_uptodate(info->chunk_root->node)) {
3555                 fprintf(stderr, "Critical roots corrupted, unable to fsck the FS\n");
3556                 return -EIO;
3557         }
3558
3559         root = info->fs_root;
3560
3561         fprintf(stderr, "checking extents\n");
3562         if (rw)
3563                 trans = btrfs_start_transaction(root, 1);
3564
3565         if (init_csum_tree) {
3566                 fprintf(stderr, "Reinit crc root\n");
3567                 ret = btrfs_fsck_reinit_root(trans, info->csum_root);
3568                 if (ret) {
3569                         fprintf(stderr, "crc root initialization failed\n");
3570                         return -EIO;
3571                 }
3572                 goto out;
3573         }
3574         ret = check_extents(trans, root, repair);
3575         if (ret)
3576                 fprintf(stderr, "Errors found in extent allocation tree\n");
3577
3578         fprintf(stderr, "checking fs roots\n");
3579         ret = check_fs_roots(root, &root_cache);
3580         if (ret)
3581                 goto out;
3582
3583         fprintf(stderr, "checking root refs\n");
3584         ret = check_root_refs(root, &root_cache);
3585 out:
3586         free_root_recs(&root_cache);
3587         if (rw) {
3588                 ret = btrfs_commit_transaction(trans, root);
3589                 if (ret)
3590                         exit(1);
3591         }
3592         close_ctree(root);
3593
3594         if (found_old_backref) { /*
3595                  * there was a disk format change when mixed
3596                  * backref was in testing tree. The old format
3597                  * existed about one week.
3598                  */
3599                 printf("\n * Found old mixed backref format. "
3600                        "The old format is not supported! *"
3601                        "\n * Please mount the FS in readonly mode, "
3602                        "backup data and re-format the FS. *\n\n");
3603                 ret = 1;
3604         }
3605         printf("found %llu bytes used err is %d\n",
3606                (unsigned long long)bytes_used, ret);
3607         printf("total csum bytes: %llu\n",(unsigned long long)total_csum_bytes);
3608         printf("total tree bytes: %llu\n",
3609                (unsigned long long)total_btree_bytes);
3610         printf("total fs tree bytes: %llu\n",
3611                (unsigned long long)total_fs_tree_bytes);
3612         printf("btree space waste bytes: %llu\n",
3613                (unsigned long long)btree_space_waste);
3614         printf("file data blocks allocated: %llu\n referenced %llu\n",
3615                 (unsigned long long)data_bytes_allocated,
3616                 (unsigned long long)data_bytes_referenced);
3617         printf("%s\n", BTRFS_BUILD_VERSION);
3618         return ret;
3619 }