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                         u64 parent = root->objectid;
 861
 862                         if (is_child_root(root, parent, location.objectid))
 863                                 add_inode_backref(root_cache, location.objectid,
 864                                                   key->objectid, key->offset,
 865                                                   namebuf, len, filetype,
 866                                                   key->type, error);
 867                 } else {
 868                         fprintf(stderr, "warning line %d\n", __LINE__);
 869                 }
 870
 871                 len = sizeof(*di) + name_len + data_len;
 872                 di = (struct btrfs_dir_item *)((char *)di + len);
 873                 cur += len;
 874         }
 875         if (key->type == BTRFS_DIR_INDEX_KEY && nritems > 1)
 876                 rec->errors |= I_ERR_DUP_DIR_INDEX;
 877
 878         return 0;
 879 }
 880
 881 static int process_inode_ref(struct extent_buffer *eb,
 882                              int slot, struct btrfs_key *key,
 883                              struct shared_node *active_node)
 884 {
 885         u32 total;
 886         u32 cur = 0;
 887         u32 len;
 888         u32 name_len;
 889         u64 index;
 890         int error;
 891         struct cache_tree *inode_cache;
 892         struct btrfs_inode_ref *ref;
 893         char namebuf[BTRFS_NAME_LEN];
 894
 895         inode_cache = &active_node->inode_cache;
 896
 897         ref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
 898         total = btrfs_item_size_nr(eb, slot);
 899         while (cur < total) {
 900                 name_len = btrfs_inode_ref_name_len(eb, ref);
 901                 index = btrfs_inode_ref_index(eb, ref);
 902                 if (name_len <= BTRFS_NAME_LEN) {
 903                         len = name_len;
 904                         error = 0;
 905                 } else {
 906                         len = BTRFS_NAME_LEN;
 907                         error = REF_ERR_NAME_TOO_LONG;
 908                 }
 909                 read_extent_buffer(eb, namebuf, (unsigned long)(ref + 1), len);
 910                 add_inode_backref(inode_cache, key->objectid, key->offset,
 911                                   index, namebuf, len, 0, key->type, error);
 912
 913                 len = sizeof(*ref) + name_len;
 914                 ref = (struct btrfs_inode_ref *)((char *)ref + len);
 915                 cur += len;
 916         }
 917         return 0;
 918 }
 919
 920 static u64 count_csum_range(struct btrfs_root *root, u64 start, u64 len)
 921 {
 922         struct btrfs_key key;
 923         struct btrfs_path path;
 924         struct extent_buffer *leaf;
 925         int ret ;
 926         size_t size;
 927         u64 found = 0;
 928         u64 csum_end;
 929         u16 csum_size = btrfs_super_csum_size(&root->fs_info->super_copy);
 930
 931         btrfs_init_path(&path);
 932
 933         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
 934         key.offset = start;
 935         key.type = BTRFS_EXTENT_CSUM_KEY;
 936
 937         ret = btrfs_search_slot(NULL, root->fs_info->csum_root,
 938                                 &key, &path, 0, 0);
 939         BUG_ON(ret < 0);
 940         if (ret > 0 && path.slots[0] > 0) {
 941                 leaf = path.nodes[0];
 942                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0] - 1);
 943                 if (key.objectid == BTRFS_EXTENT_CSUM_OBJECTID &&
 944                     key.type == BTRFS_EXTENT_CSUM_KEY)
 945                         path.slots[0]--;
 946         }
 947
 948         while (len > 0) {
 949                 leaf = path.nodes[0];
 950                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
 951                         ret = btrfs_next_leaf(root->fs_info->csum_root, &path);
 952                         BUG_ON(ret < 0);
 953                         if (ret > 0)
 954                                 break;
 955                         leaf = path.nodes[0];
 956                 }
 957
 958                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
 959                 if (key.objectid != BTRFS_EXTENT_CSUM_OBJECTID ||
 960                     key.type != BTRFS_EXTENT_CSUM_KEY)
 961                         break;
 962
 963                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
 964                 if (key.offset >= start + len)
 965                         break;
 966
 967                 if (key.offset > start)
 968                         start = key.offset;
 969
 970                 size = btrfs_item_size_nr(leaf, path.slots[0]);
 971                 csum_end = key.offset + (size / csum_size) * root->sectorsize;
 972                 if (csum_end > start) {
 973                         size = min(csum_end - start, len);
 974                         len -= size;
 975                         start += size;
 976                         found += size;
 977                 }
 978
 979                 path.slots[0]++;
 980         }
 981         btrfs_release_path(root->fs_info->csum_root, &path);
 982         return found;
 983 }
 984
 985 static int process_file_extent(struct btrfs_root *root,
 986                                 struct extent_buffer *eb,
 987                                 int slot, struct btrfs_key *key,
 988                                 struct shared_node *active_node)
 989 {
 990         struct inode_record *rec;
 991         struct btrfs_file_extent_item *fi;
 992         u64 num_bytes = 0;
 993         u64 disk_bytenr = 0;
 994         u64 extent_offset = 0;
 995         u64 mask = root->sectorsize - 1;
 996         int extent_type;
 997
 998         rec = active_node->current;
 999         BUG_ON(rec->ino != key->objectid || rec->refs > 1);
1000         rec->found_file_extent = 1;
1001
1002         if (rec->extent_start == (u64)-1) {
1003                 rec->extent_start = key->offset;
1004                 rec->extent_end = key->offset;
1005         }
1006
1007         if (rec->extent_end > key->offset)
1008                 rec->errors |= I_ERR_FILE_EXTENT_OVERLAP;
1009         else if (rec->extent_end < key->offset &&
1010                  rec->extent_end < rec->first_extent_gap)
1011                 rec->first_extent_gap = rec->extent_end;
1012
1013         fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
1014         extent_type = btrfs_file_extent_type(eb, fi);
1015
1016         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1017                 num_bytes = btrfs_file_extent_inline_len(eb, fi);
1018                 if (num_bytes == 0)
1019                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1020                 rec->found_size += num_bytes;
1021                 num_bytes = (num_bytes + mask) & ~mask;
1022         } else if (extent_type == BTRFS_FILE_EXTENT_REG ||
1023                    extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1024                 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1025                 disk_bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1026                 extent_offset = btrfs_file_extent_offset(eb, fi);
1027                 if (num_bytes == 0 || (num_bytes & mask))
1028                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1029                 if (num_bytes + extent_offset >
1030                     btrfs_file_extent_ram_bytes(eb, fi))
1031                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1032                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC &&
1033                     (btrfs_file_extent_compression(eb, fi) ||
1034                      btrfs_file_extent_encryption(eb, fi) ||
1035                      btrfs_file_extent_other_encoding(eb, fi)))
1036                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1037                 if (disk_bytenr > 0)
1038                         rec->found_size += num_bytes;
1039         } else {
1040                 rec->errors |= I_ERR_BAD_FILE_EXTENT;
1041         }
1042         rec->extent_end = key->offset + num_bytes;
1043
1044         if (disk_bytenr > 0) {
1045                 u64 found;
1046                 if (btrfs_file_extent_compression(eb, fi))
1047                         num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1048                 else
1049                         disk_bytenr += extent_offset;
1050
1051                 found = count_csum_range(root, disk_bytenr, num_bytes);
1052                 if (extent_type == BTRFS_FILE_EXTENT_REG) {
1053                         if (found > 0)
1054                                 rec->found_csum_item = 1;
1055                         if (found < num_bytes)
1056                                 rec->some_csum_missing = 1;
1057                 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1058                         if (found > 0)
1059                                 rec->errors |= I_ERR_ODD_CSUM_ITEM;
1060                 }
1061         }
1062         return 0;
1063 }
1064
1065 static int process_one_leaf(struct btrfs_root *root, struct extent_buffer *eb,
1066                             struct walk_control *wc)
1067 {
1068         struct btrfs_key key;
1069         u32 nritems;
1070         int i;
1071         int ret = 0;
1072         struct cache_tree *inode_cache;
1073         struct shared_node *active_node;
1074
1075         if (wc->root_level == wc->active_node &&
1076             btrfs_root_refs(&root->root_item) == 0)
1077                 return 0;
1078
1079         active_node = wc->nodes[wc->active_node];
1080         inode_cache = &active_node->inode_cache;
1081         nritems = btrfs_header_nritems(eb);
1082         for (i = 0; i < nritems; i++) {
1083                 btrfs_item_key_to_cpu(eb, &key, i);
1084
1085                 if (key.objectid == BTRFS_FREE_SPACE_OBJECTID)
1086                         continue;
1087
1088                 if (active_node->current == NULL ||
1089                     active_node->current->ino < key.objectid) {
1090                         if (active_node->current) {
1091                                 active_node->current->checked = 1;
1092                                 maybe_free_inode_rec(inode_cache,
1093                                                      active_node->current);
1094                         }
1095                         active_node->current = get_inode_rec(inode_cache,
1096                                                              key.objectid, 1);
1097                 }
1098                 switch (key.type) {
1099                 case BTRFS_DIR_ITEM_KEY:
1100                 case BTRFS_DIR_INDEX_KEY:
1101                         ret = process_dir_item(root, eb, i, &key, active_node);
1102                         break;
1103                 case BTRFS_INODE_REF_KEY:
1104                         ret = process_inode_ref(eb, i, &key, active_node);
1105                         break;
1106                 case BTRFS_INODE_ITEM_KEY:
1107                         ret = process_inode_item(eb, i, &key, active_node);
1108                         break;
1109                 case BTRFS_EXTENT_DATA_KEY:
1110                         ret = process_file_extent(root, eb, i, &key,
1111                                                   active_node);
1112                         break;
1113                 default:
1114                         break;
1115                 };
1116         }
1117         return ret;
1118 }
1119
1120 static void reada_walk_down(struct btrfs_root *root,
1121                             struct extent_buffer *node, int slot)
1122 {
1123         u64 bytenr;
1124         u64 ptr_gen;
1125         u32 nritems;
1126         u32 blocksize;
1127         int i;
1128         int ret;
1129         int level;
1130
1131         level = btrfs_header_level(node);
1132         if (level != 1)
1133                 return;
1134
1135         nritems = btrfs_header_nritems(node);
1136         blocksize = btrfs_level_size(root, level - 1);
1137         for (i = slot; i < nritems; i++) {
1138                 bytenr = btrfs_node_blockptr(node, i);
1139                 ptr_gen = btrfs_node_ptr_generation(node, i);
1140                 ret = readahead_tree_block(root, bytenr, blocksize, ptr_gen);
1141                 if (ret)
1142                         break;
1143         }
1144 }
1145
1146 static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
1147                           struct walk_control *wc, int *level)
1148 {
1149         u64 bytenr;
1150         u64 ptr_gen;
1151         struct extent_buffer *next;
1152         struct extent_buffer *cur;
1153         u32 blocksize;
1154         int ret;
1155         u64 refs;
1156
1157         WARN_ON(*level < 0);
1158         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1159         ret = btrfs_lookup_extent_info(NULL, root,
1160                                        path->nodes[*level]->start,
1161                                        path->nodes[*level]->len, &refs, NULL);
1162         if (ret < 0)
1163                 goto out;
1164
1165         if (refs > 1) {
1166                 ret = enter_shared_node(root, path->nodes[*level]->start,
1167                                         refs, wc, *level);
1168                 if (ret > 0)
1169                         goto out;
1170         }
1171
1172         while (*level >= 0) {
1173                 WARN_ON(*level < 0);
1174                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1175                 cur = path->nodes[*level];
1176
1177                 if (btrfs_header_level(cur) != *level)
1178                         WARN_ON(1);
1179
1180                 if (path->slots[*level] >= btrfs_header_nritems(cur))
1181                         break;
1182                 if (*level == 0) {
1183                         ret = process_one_leaf(root, cur, wc);
1184                         break;
1185                 }
1186                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1187                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1188                 blocksize = btrfs_level_size(root, *level - 1);
1189                 ret = btrfs_lookup_extent_info(NULL, root, bytenr, blocksize,
1190                                                &refs, NULL);
1191                 if (ret < 0)
1192                         refs = 0;
1193
1194                 if (refs > 1) {
1195                         ret = enter_shared_node(root, bytenr, refs,
1196                                                 wc, *level - 1);
1197                         if (ret > 0) {
1198                                 path->slots[*level]++;
1199                                 continue;
1200                         }
1201                 }
1202
1203                 next = btrfs_find_tree_block(root, bytenr, blocksize);
1204                 if (!next || !btrfs_buffer_uptodate(next, ptr_gen)) {
1205                         free_extent_buffer(next);
1206                         reada_walk_down(root, cur, path->slots[*level]);
1207                         next = read_tree_block(root, bytenr, blocksize,
1208                                                ptr_gen);
1209                 }
1210
1211                 *level = *level - 1;
1212                 free_extent_buffer(path->nodes[*level]);
1213                 path->nodes[*level] = next;
1214                 path->slots[*level] = 0;
1215         }
1216 out:
1217         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
1218         return 0;
1219 }
1220
1221 static int walk_up_tree(struct btrfs_root *root, struct btrfs_path *path,
1222                         struct walk_control *wc, int *level)
1223 {
1224         int i;
1225         struct extent_buffer *leaf;
1226
1227         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
1228                 leaf = path->nodes[i];
1229                 if (path->slots[i] + 1 < btrfs_header_nritems(leaf)) {
1230                         path->slots[i]++;
1231                         *level = i;
1232                         return 0;
1233                 } else {
1234                         free_extent_buffer(path->nodes[*level]);
1235                         path->nodes[*level] = NULL;
1236                         BUG_ON(*level > wc->active_node);
1237                         if (*level == wc->active_node)
1238                                 leave_shared_node(root, wc, *level);
1239                         *level = i + 1;
1240                 }
1241         }
1242         return 1;
1243 }
1244
1245 static int check_root_dir(struct inode_record *rec)
1246 {
1247         struct inode_backref *backref;
1248         int ret = -1;
1249
1250         if (!rec->found_inode_item || rec->errors)
1251                 goto out;
1252         if (rec->nlink != 1 || rec->found_link != 0)
1253                 goto out;
1254         if (list_empty(&rec->backrefs))
1255                 goto out;
1256         backref = list_entry(rec->backrefs.next, struct inode_backref, list);
1257         if (!backref->found_inode_ref)
1258                 goto out;
1259         if (backref->index != 0 || backref->namelen != 2 ||
1260             memcmp(backref->name, "..", 2))
1261                 goto out;
1262         if (backref->found_dir_index || backref->found_dir_item)
1263                 goto out;
1264         ret = 0;
1265 out:
1266         return ret;
1267 }
1268
1269 static int check_inode_recs(struct btrfs_root *root,
1270                             struct cache_tree *inode_cache)
1271 {
1272         struct cache_extent *cache;
1273         struct ptr_node *node;
1274         struct inode_record *rec;
1275         struct inode_backref *backref;
1276         int ret;
1277         u64 error = 0;
1278         u64 root_dirid = btrfs_root_dirid(&root->root_item);
1279
1280         if (btrfs_root_refs(&root->root_item) == 0) {
1281                 if (!cache_tree_empty(inode_cache))
1282                         fprintf(stderr, "warning line %d\n", __LINE__);
1283                 return 0;
1284         }
1285
1286         rec = get_inode_rec(inode_cache, root_dirid, 0);
1287         if (rec) {
1288                 ret = check_root_dir(rec);
1289                 if (ret) {
1290                         fprintf(stderr, "root %llu root dir %llu error\n",
1291                                 (unsigned long long)root->root_key.objectid,
1292                                 (unsigned long long)root_dirid);
1293                         error++;
1294                 }
1295         } else {
1296                 fprintf(stderr, "root %llu root dir %llu not found\n",
1297                         (unsigned long long)root->root_key.objectid,
1298                         (unsigned long long)root_dirid);
1299         }
1300
1301         while (1) {
1302                 cache = find_first_cache_extent(inode_cache, 0);
1303                 if (!cache)
1304                         break;
1305                 node = container_of(cache, struct ptr_node, cache);
1306                 rec = node->data;
1307                 remove_cache_extent(inode_cache, &node->cache);
1308                 free(node);
1309                 if (rec->ino == root_dirid ||
1310                     rec->ino == BTRFS_ORPHAN_OBJECTID) {
1311                         free_inode_rec(rec);
1312                         continue;
1313                 }
1314
1315                 if (rec->errors & I_ERR_NO_ORPHAN_ITEM) {
1316                         ret = check_orphan_item(root, rec->ino);
1317                         if (ret == 0)
1318                                 rec->errors &= ~I_ERR_NO_ORPHAN_ITEM;
1319                         if (can_free_inode_rec(rec)) {
1320                                 free_inode_rec(rec);
1321                                 continue;
1322                         }
1323                 }
1324
1325                 error++;
1326                 if (!rec->found_inode_item)
1327                         rec->errors |= I_ERR_NO_INODE_ITEM;
1328                 if (rec->found_link != rec->nlink)
1329                         rec->errors |= I_ERR_LINK_COUNT_WRONG;
1330                 fprintf(stderr, "root %llu inode %llu errors %x\n",
1331                         (unsigned long long) root->root_key.objectid,
1332                         (unsigned long long) rec->ino, rec->errors);
1333                 list_for_each_entry(backref, &rec->backrefs, list) {
1334                         if (!backref->found_dir_item)
1335                                 backref->errors |= REF_ERR_NO_DIR_ITEM;
1336                         if (!backref->found_dir_index)
1337                                 backref->errors |= REF_ERR_NO_DIR_INDEX;
1338                         if (!backref->found_inode_ref)
1339                                 backref->errors |= REF_ERR_NO_INODE_REF;
1340                         fprintf(stderr, "\tunresolved ref dir %llu index %llu"
1341                                 " namelen %u name %s filetype %d error %x\n",
1342                                 (unsigned long long)backref->dir,
1343                                 (unsigned long long)backref->index,
1344                                 backref->namelen, backref->name,
1345                                 backref->filetype, backref->errors);
1346                 }
1347                 free_inode_rec(rec);
1348         }
1349         return (error > 0) ? -1 : 0;
1350 }
1351
1352 static struct root_record *get_root_rec(struct cache_tree *root_cache,
1353                                         u64 objectid)
1354 {
1355         struct cache_extent *cache;
1356         struct root_record *rec = NULL;
1357         int ret;
1358
1359         cache = find_cache_extent(root_cache, objectid, 1);
1360         if (cache) {
1361                 rec = container_of(cache, struct root_record, cache);
1362         } else {
1363                 rec = calloc(1, sizeof(*rec));
1364                 rec->objectid = objectid;
1365                 INIT_LIST_HEAD(&rec->backrefs);
1366                 rec->cache.start = objectid;
1367                 rec->cache.size = 1;
1368
1369                 ret = insert_existing_cache_extent(root_cache, &rec->cache);
1370                 BUG_ON(ret);
1371         }
1372         return rec;
1373 }
1374
1375 static struct root_backref *get_root_backref(struct root_record *rec,
1376                                              u64 ref_root, u64 dir, u64 index,
1377                                              const char *name, int namelen)
1378 {
1379         struct root_backref *backref;
1380
1381         list_for_each_entry(backref, &rec->backrefs, list) {
1382                 if (backref->ref_root != ref_root || backref->dir != dir ||
1383                     backref->namelen != namelen)
1384                         continue;
1385                 if (memcmp(name, backref->name, namelen))
1386                         continue;
1387                 return backref;
1388         }
1389
1390         backref = malloc(sizeof(*backref) + namelen + 1);
1391         memset(backref, 0, sizeof(*backref));
1392         backref->ref_root = ref_root;
1393         backref->dir = dir;
1394         backref->index = index;
1395         backref->namelen = namelen;
1396         memcpy(backref->name, name, namelen);
1397         backref->name[namelen] = '\0';
1398         list_add_tail(&backref->list, &rec->backrefs);
1399         return backref;
1400 }
1401
1402 static void free_root_recs(struct cache_tree *root_cache)
1403 {
1404         struct cache_extent *cache;
1405         struct root_record *rec;
1406         struct root_backref *backref;
1407
1408         while (1) {
1409                 cache = find_first_cache_extent(root_cache, 0);
1410                 if (!cache)
1411                         break;
1412                 rec = container_of(cache, struct root_record, cache);
1413                 remove_cache_extent(root_cache, &rec->cache);
1414
1415                 while (!list_empty(&rec->backrefs)) {
1416                         backref = list_entry(rec->backrefs.next,
1417                                              struct root_backref, list);
1418                         list_del(&backref->list);
1419                         free(backref);
1420                 }
1421                 kfree(rec);
1422         }
1423 }
1424
1425 static int add_root_backref(struct cache_tree *root_cache,
1426                             u64 root_id, u64 ref_root, u64 dir, u64 index,
1427                             const char *name, int namelen,
1428                             int item_type, int errors)
1429 {
1430         struct root_record *rec;
1431         struct root_backref *backref;
1432
1433         rec = get_root_rec(root_cache, root_id);
1434         backref = get_root_backref(rec, ref_root, dir, index, name, namelen);
1435
1436         backref->errors |= errors;
1437
1438         if (item_type != BTRFS_DIR_ITEM_KEY) {
1439                 if (backref->found_dir_index || backref->found_back_ref ||
1440                     backref->found_forward_ref) {
1441                         if (backref->index != index)
1442                                 backref->errors |= REF_ERR_INDEX_UNMATCH;
1443                 } else {
1444                         backref->index = index;
1445                 }
1446         }
1447
1448         if (item_type == BTRFS_DIR_ITEM_KEY) {
1449                 backref->found_dir_item = 1;
1450                 backref->reachable = 1;
1451                 rec->found_ref++;
1452         } else if (item_type == BTRFS_DIR_INDEX_KEY) {
1453                 backref->found_dir_index = 1;
1454         } else if (item_type == BTRFS_ROOT_REF_KEY) {
1455                 if (backref->found_forward_ref)
1456                         backref->errors |= REF_ERR_DUP_ROOT_REF;
1457                 backref->found_forward_ref = 1;
1458         } else if (item_type == BTRFS_ROOT_BACKREF_KEY) {
1459                 if (backref->found_back_ref)
1460                         backref->errors |= REF_ERR_DUP_ROOT_BACKREF;
1461                 backref->found_back_ref = 1;
1462         } else {
1463                 BUG_ON(1);
1464         }
1465
1466         return 0;
1467 }
1468
1469 static int merge_root_recs(struct btrfs_root *root,
1470                            struct cache_tree *src_cache,
1471                            struct cache_tree *dst_cache)
1472 {
1473         struct cache_extent *cache;
1474         struct ptr_node *node;
1475         struct inode_record *rec;
1476         struct inode_backref *backref;
1477
1478         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
1479                 free_inode_recs(src_cache);
1480                 return 0;
1481         }
1482
1483         while (1) {
1484                 cache = find_first_cache_extent(src_cache, 0);
1485                 if (!cache)
1486                         break;
1487                 node = container_of(cache, struct ptr_node, cache);
1488                 rec = node->data;
1489                 remove_cache_extent(src_cache, &node->cache);
1490                 free(node);
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                 free_inode_rec(rec);
1508         }
1509         return 0;
1510 }
1511
1512 static int check_root_refs(struct btrfs_root *root,
1513                            struct cache_tree *root_cache)
1514 {
1515         struct root_record *rec;
1516         struct root_record *ref_root;
1517         struct root_backref *backref;
1518         struct cache_extent *cache;
1519         int loop = 1;
1520         int ret;
1521         int error;
1522         int errors = 0;
1523
1524         rec = get_root_rec(root_cache, BTRFS_FS_TREE_OBJECTID);
1525         rec->found_ref = 1;
1526
1527         /* fixme: this can not detect circular references */
1528         while (loop) {
1529                 loop = 0;
1530                 cache = find_first_cache_extent(root_cache, 0);
1531                 while (1) {
1532                         if (!cache)
1533                                 break;
1534                         rec = container_of(cache, struct root_record, cache);
1535                         cache = next_cache_extent(cache);
1536
1537                         if (rec->found_ref == 0)
1538                                 continue;
1539
1540                         list_for_each_entry(backref, &rec->backrefs, list) {
1541                                 if (!backref->reachable)
1542                                         continue;
1543
1544                                 ref_root = get_root_rec(root_cache,
1545                                                         backref->ref_root);
1546                                 if (ref_root->found_ref > 0)
1547                                         continue;
1548
1549                                 backref->reachable = 0;
1550                                 rec->found_ref--;
1551                                 if (rec->found_ref == 0)
1552                                         loop = 1;
1553                         }
1554                 }
1555         }
1556
1557         cache = find_first_cache_extent(root_cache, 0);
1558         while (1) {
1559                 if (!cache)
1560                         break;
1561                 rec = container_of(cache, struct root_record, cache);
1562                 cache = next_cache_extent(cache);
1563
1564                 if (rec->found_ref == 0 &&
1565                     rec->objectid >= BTRFS_FIRST_FREE_OBJECTID &&
1566                     rec->objectid <= BTRFS_LAST_FREE_OBJECTID) {
1567                         ret = check_orphan_item(root->fs_info->tree_root,
1568                                                 rec->objectid);
1569                         if (ret == 0)
1570                                 continue;
1571                         errors++;
1572                         fprintf(stderr, "fs tree %llu not referenced\n",
1573                                 (unsigned long long)rec->objectid);
1574                 }
1575
1576                 error = 0;
1577                 if (rec->found_ref > 0 && !rec->found_root_item)
1578                         error = 1;
1579                 list_for_each_entry(backref, &rec->backrefs, list) {
1580                         if (!backref->found_dir_item)
1581                                 backref->errors |= REF_ERR_NO_DIR_ITEM;
1582                         if (!backref->found_dir_index)
1583                                 backref->errors |= REF_ERR_NO_DIR_INDEX;
1584                         if (!backref->found_back_ref)
1585                                 backref->errors |= REF_ERR_NO_ROOT_BACKREF;
1586                         if (!backref->found_forward_ref)
1587                                 backref->errors |= REF_ERR_NO_ROOT_REF;
1588                         if (backref->reachable && backref->errors)
1589                                 error = 1;
1590                 }
1591                 if (!error)
1592                         continue;
1593
1594                 errors++;
1595                 fprintf(stderr, "fs tree %llu refs %u %s\n",
1596                         (unsigned long long)rec->objectid, rec->found_ref,
1597                          rec->found_root_item ? "" : "not found");
1598
1599                 list_for_each_entry(backref, &rec->backrefs, list) {
1600                         if (!backref->reachable)
1601                                 continue;
1602                         if (!backref->errors && rec->found_root_item)
1603                                 continue;
1604                         fprintf(stderr, "\tunresolved ref root %llu dir %llu"
1605                                 " index %llu namelen %u name %s error %x\n",
1606                                 (unsigned long long)backref->ref_root,
1607                                 (unsigned long long)backref->dir,
1608                                 (unsigned long long)backref->index,
1609                                 backref->namelen, backref->name,
1610                                 backref->errors);
1611                 }
1612         }
1613         return errors > 0 ? 1 : 0;
1614 }
1615
1616 static int process_root_ref(struct extent_buffer *eb, int slot,
1617                             struct btrfs_key *key,
1618                             struct cache_tree *root_cache)
1619 {
1620         u64 dirid;
1621         u64 index;
1622         u32 len;
1623         u32 name_len;
1624         struct btrfs_root_ref *ref;
1625         char namebuf[BTRFS_NAME_LEN];
1626         int error;
1627
1628         ref = btrfs_item_ptr(eb, slot, struct btrfs_root_ref);
1629
1630         dirid = btrfs_root_ref_dirid(eb, ref);
1631         index = btrfs_root_ref_sequence(eb, ref);
1632         name_len = btrfs_root_ref_name_len(eb, ref);
1633
1634         if (name_len <= BTRFS_NAME_LEN) {
1635                 len = name_len;
1636                 error = 0;
1637         } else {
1638                 len = BTRFS_NAME_LEN;
1639                 error = REF_ERR_NAME_TOO_LONG;
1640         }
1641         read_extent_buffer(eb, namebuf, (unsigned long)(ref + 1), len);
1642
1643         if (key->type == BTRFS_ROOT_REF_KEY) {
1644                 add_root_backref(root_cache, key->offset, key->objectid, dirid,
1645                                  index, namebuf, len, key->type, error);
1646         } else {
1647                 add_root_backref(root_cache, key->objectid, key->offset, dirid,
1648                                  index, namebuf, len, key->type, error);
1649         }
1650         return 0;
1651 }
1652
1653 static int check_fs_root(struct btrfs_root *root,
1654                          struct cache_tree *root_cache,
1655                          struct walk_control *wc)
1656 {
1657         int ret = 0;
1658         int wret;
1659         int level;
1660         struct btrfs_path path;
1661         struct shared_node root_node;
1662         struct root_record *rec;
1663         struct btrfs_root_item *root_item = &root->root_item;
1664
1665         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1666                 rec = get_root_rec(root_cache, root->root_key.objectid);
1667                 if (btrfs_root_refs(root_item) > 0)
1668                         rec->found_root_item = 1;
1669         }
1670
1671         btrfs_init_path(&path);
1672         memset(&root_node, 0, sizeof(root_node));
1673         cache_tree_init(&root_node.root_cache);
1674         cache_tree_init(&root_node.inode_cache);
1675
1676         level = btrfs_header_level(root->node);
1677         memset(wc->nodes, 0, sizeof(wc->nodes));
1678         wc->nodes[level] = &root_node;
1679         wc->active_node = level;
1680         wc->root_level = level;
1681
1682         if (btrfs_root_refs(root_item) > 0 ||
1683             btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1684                 path.nodes[level] = root->node;
1685                 extent_buffer_get(root->node);
1686                 path.slots[level] = 0;
1687         } else {
1688                 struct btrfs_key key;
1689                 struct btrfs_disk_key found_key;
1690
1691                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1692                 level = root_item->drop_level;
1693                 path.lowest_level = level;
1694                 wret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
1695                 BUG_ON(wret < 0);
1696                 btrfs_node_key(path.nodes[level], &found_key,
1697                                 path.slots[level]);
1698                 WARN_ON(memcmp(&found_key, &root_item->drop_progress,
1699                                         sizeof(found_key)));
1700         }
1701
1702         while (1) {
1703                 wret = walk_down_tree(root, &path, wc, &level);
1704                 if (wret < 0)
1705                         ret = wret;
1706                 if (wret != 0)
1707                         break;
1708
1709                 wret = walk_up_tree(root, &path, wc, &level);
1710                 if (wret < 0)
1711                         ret = wret;
1712                 if (wret != 0)
1713                         break;
1714         }
1715         btrfs_release_path(root, &path);
1716
1717         merge_root_recs(root, &root_node.root_cache, root_cache);
1718
1719         if (root_node.current) {
1720                 root_node.current->checked = 1;
1721                 maybe_free_inode_rec(&root_node.inode_cache,
1722                                 root_node.current);
1723         }
1724
1725         ret = check_inode_recs(root, &root_node.inode_cache);
1726         return ret;
1727 }
1728
1729 static int fs_root_objectid(u64 objectid)
1730 {
1731         if (objectid == BTRFS_FS_TREE_OBJECTID ||
1732             objectid == BTRFS_TREE_RELOC_OBJECTID ||
1733             objectid == BTRFS_DATA_RELOC_TREE_OBJECTID ||
1734             (objectid >= BTRFS_FIRST_FREE_OBJECTID &&
1735              objectid <= BTRFS_LAST_FREE_OBJECTID))
1736                 return 1;
1737         return 0;
1738 }
1739
1740 static int check_fs_roots(struct btrfs_root *root,
1741                           struct cache_tree *root_cache)
1742 {
1743         struct btrfs_path path;
1744         struct btrfs_key key;
1745         struct walk_control wc;
1746         struct extent_buffer *leaf;
1747         struct btrfs_root *tmp_root;
1748         struct btrfs_root *tree_root = root->fs_info->tree_root;
1749         int ret;
1750         int err = 0;
1751
1752         memset(&wc, 0, sizeof(wc));
1753         cache_tree_init(&wc.shared);
1754         btrfs_init_path(&path);
1755
1756         key.offset = 0;
1757         key.objectid = 0;
1758         key.type = BTRFS_ROOT_ITEM_KEY;
1759         ret = btrfs_search_slot(NULL, tree_root, &key, &path, 0, 0);
1760         BUG_ON(ret < 0);
1761         while (1) {
1762                 leaf = path.nodes[0];
1763                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1764                         ret = btrfs_next_leaf(tree_root, &path);
1765                         if (ret != 0)
1766                                 break;
1767                         leaf = path.nodes[0];
1768                 }
1769                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1770                 if (key.type == BTRFS_ROOT_ITEM_KEY &&
1771                     fs_root_objectid(key.objectid)) {
1772                         tmp_root = btrfs_read_fs_root_no_cache(root->fs_info,
1773                                                                &key);
1774                         ret = check_fs_root(tmp_root, root_cache, &wc);
1775                         if (ret)
1776                                 err = 1;
1777                         btrfs_free_fs_root(root->fs_info, tmp_root);
1778                 } else if (key.type == BTRFS_ROOT_REF_KEY ||
1779                            key.type == BTRFS_ROOT_BACKREF_KEY) {
1780                         process_root_ref(leaf, path.slots[0], &key,
1781                                          root_cache);
1782                 }
1783                 path.slots[0]++;
1784         }
1785         btrfs_release_path(tree_root, &path);
1786
1787         if (!cache_tree_empty(&wc.shared))
1788                 fprintf(stderr, "warning line %d\n", __LINE__);
1789
1790         return err;
1791 }
1792
1793 static int all_backpointers_checked(struct extent_record *rec, int print_errs)
1794 {
1795         struct list_head *cur = rec->backrefs.next;
1796         struct extent_backref *back;
1797         struct tree_backref *tback;
1798         struct data_backref *dback;
1799         u64 found = 0;
1800         int err = 0;
1801
1802         while(cur != &rec->backrefs) {
1803                 back = list_entry(cur, struct extent_backref, list);
1804                 cur = cur->next;
1805                 if (!back->found_extent_tree) {
1806                         err = 1;
1807                         if (!print_errs)
1808                                 goto out;
1809                         if (back->is_data) {
1810                                 dback = (struct data_backref *)back;
1811                                 fprintf(stderr, "Backref %llu %s %llu"
1812                                         " owner %llu offset %llu num_refs %lu"
1813                                         " not found in extent tree\n",
1814                                         (unsigned long long)rec->start,
1815                                         back->full_backref ?
1816                                         "parent" : "root",
1817                                         back->full_backref ?
1818                                         (unsigned long long)dback->parent:
1819                                         (unsigned long long)dback->root,
1820                                         (unsigned long long)dback->owner,
1821                                         (unsigned long long)dback->offset,
1822                                         (unsigned long)dback->num_refs);
1823                         } else {
1824                                 tback = (struct tree_backref *)back;
1825                                 fprintf(stderr, "Backref %llu parent %llu"
1826                                         " root %llu not found in extent tree\n",
1827                                         (unsigned long long)rec->start,
1828                                         (unsigned long long)tback->parent,
1829                                         (unsigned long long)tback->root);
1830                         }
1831                 }
1832                 if (!back->is_data && !back->found_ref) {
1833                         err = 1;
1834                         if (!print_errs)
1835                                 goto out;
1836                         tback = (struct tree_backref *)back;
1837                         fprintf(stderr, "Backref %llu %s %llu not referenced back %p\n",
1838                                 (unsigned long long)rec->start,
1839                                 back->full_backref ? "parent" : "root",
1840                                 back->full_backref ?
1841                                 (unsigned long long)tback->parent :
1842                                 (unsigned long long)tback->root, back);
1843                 }
1844                 if (back->is_data) {
1845                         dback = (struct data_backref *)back;
1846                         if (dback->found_ref != dback->num_refs) {
1847                                 err = 1;
1848                                 if (!print_errs)
1849                                         goto out;
1850                                 fprintf(stderr, "Incorrect local backref count"
1851                                         " on %llu %s %llu owner %llu"
1852                                         " offset %llu found %u wanted %u back %p\n",
1853                                         (unsigned long long)rec->start,
1854                                         back->full_backref ?
1855                                         "parent" : "root",
1856                                         back->full_backref ?
1857                                         (unsigned long long)dback->parent:
1858                                         (unsigned long long)dback->root,
1859                                         (unsigned long long)dback->owner,
1860                                         (unsigned long long)dback->offset,
1861                                         dback->found_ref, dback->num_refs, back);
1862                         }
1863                 }
1864                 if (!back->is_data) {
1865                         found += 1;
1866                 } else {
1867                         dback = (struct data_backref *)back;
1868                         found += dback->found_ref;
1869                 }
1870         }
1871         if (found != rec->refs) {
1872                 err = 1;
1873                 if (!print_errs)
1874                         goto out;
1875                 fprintf(stderr, "Incorrect global backref count "
1876                         "on %llu found %llu wanted %llu\n",
1877                         (unsigned long long)rec->start,
1878                         (unsigned long long)found,
1879                         (unsigned long long)rec->refs);
1880         }
1881 out:
1882         return err;
1883 }
1884
1885 static int free_all_extent_backrefs(struct extent_record *rec)
1886 {
1887         struct extent_backref *back;
1888         struct list_head *cur;
1889         while (!list_empty(&rec->backrefs)) {
1890                 cur = rec->backrefs.next;
1891                 back = list_entry(cur, struct extent_backref, list);
1892                 list_del(cur);
1893                 free(back);
1894         }
1895         return 0;
1896 }
1897
1898 static int maybe_free_extent_rec(struct cache_tree *extent_cache,
1899                                  struct extent_record *rec)
1900 {
1901         if (rec->content_checked && rec->owner_ref_checked &&
1902             rec->extent_item_refs == rec->refs && rec->refs > 0 &&
1903             !all_backpointers_checked(rec, 0)) {
1904                 remove_cache_extent(extent_cache, &rec->cache);
1905                 free_all_extent_backrefs(rec);
1906                 free(rec);
1907         }
1908         return 0;
1909 }
1910
1911 static int check_owner_ref(struct btrfs_root *root,
1912                             struct extent_record *rec,
1913                             struct extent_buffer *buf)
1914 {
1915         struct extent_backref *node;
1916         struct tree_backref *back;
1917         struct btrfs_root *ref_root;
1918         struct btrfs_key key;
1919         struct btrfs_path path;
1920         int level;
1921         int found = 0;
1922
1923         list_for_each_entry(node, &rec->backrefs, list) {
1924                 if (node->is_data)
1925                         continue;
1926                 if (!node->found_ref)
1927                         continue;
1928                 if (node->full_backref)
1929                         continue;
1930                 back = (struct tree_backref *)node;
1931                 if (btrfs_header_owner(buf) == back->root)
1932                         return 0;
1933         }
1934         BUG_ON(rec->is_root);
1935
1936         /* try to find the block by search corresponding fs tree */
1937         key.objectid = btrfs_header_owner(buf);
1938         key.type = BTRFS_ROOT_ITEM_KEY;
1939         key.offset = (u64)-1;
1940
1941         ref_root = btrfs_read_fs_root(root->fs_info, &key);
1942         BUG_ON(IS_ERR(ref_root));
1943
1944         level = btrfs_header_level(buf);
1945         if (level == 0)
1946                 btrfs_item_key_to_cpu(buf, &key, 0);
1947         else
1948                 btrfs_node_key_to_cpu(buf, &key, 0);
1949
1950         btrfs_init_path(&path);
1951         path.lowest_level = level + 1;
1952         btrfs_search_slot(NULL, ref_root, &key, &path, 0, 0);
1953
1954         if (buf->start == btrfs_node_blockptr(path.nodes[level + 1],
1955                                               path.slots[level + 1]))
1956                 rec->owner_ref_checked = 1;
1957
1958         btrfs_release_path(ref_root, &path);
1959         return found ? 0 : 1;
1960 }
1961
1962 static int is_extent_tree_record(struct extent_record *rec)
1963 {
1964         struct list_head *cur = rec->backrefs.next;
1965         struct extent_backref *node;
1966         struct tree_backref *back;
1967         int is_extent = 0;
1968
1969         while(cur != &rec->backrefs) {
1970                 node = list_entry(cur, struct extent_backref, list);
1971                 cur = cur->next;
1972                 if (node->is_data)
1973                         return 0;
1974                 back = (struct tree_backref *)node;
1975                 if (node->full_backref)
1976                         return 0;
1977                 if (back->root == BTRFS_EXTENT_TREE_OBJECTID)
1978                         is_extent = 1;
1979         }
1980         return is_extent;
1981 }
1982
1983
1984 static int record_bad_block_io(struct btrfs_fs_info *info,
1985                                struct cache_tree *extent_cache,
1986                                u64 start, u64 len)
1987 {
1988         struct extent_record *rec;
1989         struct cache_extent *cache;
1990         struct btrfs_key key;
1991
1992         cache = find_cache_extent(extent_cache, start, len);
1993         if (!cache)
1994                 return 0;
1995
1996         rec = container_of(cache, struct extent_record, cache);
1997         if (!is_extent_tree_record(rec))
1998                 return 0;
1999
2000         btrfs_disk_key_to_cpu(&key, &rec->parent_key);
2001         return btrfs_add_corrupt_extent_record(info, &key, start, len, 0);
2002 }
2003
2004 static int check_block(struct btrfs_root *root,
2005                        struct cache_tree *extent_cache,
2006                        struct extent_buffer *buf, u64 flags)
2007 {
2008         struct extent_record *rec;
2009         struct cache_extent *cache;
2010         struct btrfs_key key;
2011         int ret = 1;
2012         int level;
2013
2014         cache = find_cache_extent(extent_cache, buf->start, buf->len);
2015         if (!cache)
2016                 return 1;
2017         rec = container_of(cache, struct extent_record, cache);
2018         rec->generation = btrfs_header_generation(buf);
2019
2020         level = btrfs_header_level(buf);
2021         if (btrfs_header_nritems(buf) > 0) {
2022
2023                 if (level == 0)
2024                         btrfs_item_key_to_cpu(buf, &key, 0);
2025                 else
2026                         btrfs_node_key_to_cpu(buf, &key, 0);
2027
2028                 rec->info_objectid = key.objectid;
2029         }
2030         rec->info_level = level;
2031
2032         if (btrfs_is_leaf(buf))
2033                 ret = btrfs_check_leaf(root, &rec->parent_key, buf);
2034         else
2035                 ret = btrfs_check_node(root, &rec->parent_key, buf);
2036
2037         if (ret) {
2038                 fprintf(stderr, "bad block %llu\n",
2039                         (unsigned long long)buf->start);
2040         } else {
2041                 rec->content_checked = 1;
2042                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
2043                         rec->owner_ref_checked = 1;
2044                 else {
2045                         ret = check_owner_ref(root, rec, buf);
2046                         if (!ret)
2047                                 rec->owner_ref_checked = 1;
2048                 }
2049         }
2050         if (!ret)
2051                 maybe_free_extent_rec(extent_cache, rec);
2052         return ret;
2053 }
2054
2055 static struct tree_backref *find_tree_backref(struct extent_record *rec,
2056                                                 u64 parent, u64 root)
2057 {
2058         struct list_head *cur = rec->backrefs.next;
2059         struct extent_backref *node;
2060         struct tree_backref *back;
2061
2062         while(cur != &rec->backrefs) {
2063                 node = list_entry(cur, struct extent_backref, list);
2064                 cur = cur->next;
2065                 if (node->is_data)
2066                         continue;
2067                 back = (struct tree_backref *)node;
2068                 if (parent > 0) {
2069                         if (!node->full_backref)
2070                                 continue;
2071                         if (parent == back->parent)
2072                                 return back;
2073                 } else {
2074                         if (node->full_backref)
2075                                 continue;
2076                         if (back->root == root)
2077                                 return back;
2078                 }
2079         }
2080         return NULL;
2081 }
2082
2083 static struct tree_backref *alloc_tree_backref(struct extent_record *rec,
2084                                                 u64 parent, u64 root)
2085 {
2086         struct tree_backref *ref = malloc(sizeof(*ref));
2087         memset(&ref->node, 0, sizeof(ref->node));
2088         if (parent > 0) {
2089                 ref->parent = parent;
2090                 ref->node.full_backref = 1;
2091         } else {
2092                 ref->root = root;
2093                 ref->node.full_backref = 0;
2094         }
2095         list_add_tail(&ref->node.list, &rec->backrefs);
2096
2097         return ref;
2098 }
2099
2100 static struct data_backref *find_data_backref(struct extent_record *rec,
2101                                                 u64 parent, u64 root,
2102                                                 u64 owner, u64 offset)
2103 {
2104         struct list_head *cur = rec->backrefs.next;
2105         struct extent_backref *node;
2106         struct data_backref *back;
2107
2108         while(cur != &rec->backrefs) {
2109                 node = list_entry(cur, struct extent_backref, list);
2110                 cur = cur->next;
2111                 if (!node->is_data)
2112                         continue;
2113                 back = (struct data_backref *)node;
2114                 if (parent > 0) {
2115                         if (!node->full_backref)
2116                                 continue;
2117                         if (parent == back->parent)
2118                                 return back;
2119                 } else {
2120                         if (node->full_backref)
2121                                 continue;
2122                         if (back->root == root && back->owner == owner &&
2123                             back->offset == offset)
2124                                 return back;
2125                 }
2126         }
2127         return NULL;
2128 }
2129
2130 static struct data_backref *alloc_data_backref(struct extent_record *rec,
2131                                                 u64 parent, u64 root,
2132                                                 u64 owner, u64 offset,
2133                                                 u64 max_size)
2134 {
2135         struct data_backref *ref = malloc(sizeof(*ref));
2136         memset(&ref->node, 0, sizeof(ref->node));
2137         ref->node.is_data = 1;
2138
2139         if (parent > 0) {
2140                 ref->parent = parent;
2141                 ref->owner = 0;
2142                 ref->offset = 0;
2143                 ref->node.full_backref = 1;
2144         } else {
2145                 ref->root = root;
2146                 ref->owner = owner;
2147                 ref->offset = offset;
2148                 ref->node.full_backref = 0;
2149         }
2150         ref->found_ref = 0;
2151         ref->num_refs = 0;
2152         list_add_tail(&ref->node.list, &rec->backrefs);
2153         if (max_size > rec->max_size)
2154                 rec->max_size = max_size;
2155         return ref;
2156 }
2157
2158 static int add_extent_rec(struct cache_tree *extent_cache,
2159                           struct btrfs_key *parent_key,
2160                           u64 start, u64 nr, u64 extent_item_refs,
2161                           int is_root, int inc_ref, int set_checked,
2162                           u64 max_size)
2163 {
2164         struct extent_record *rec;
2165         struct cache_extent *cache;
2166         int ret = 0;
2167
2168         cache = find_cache_extent(extent_cache, start, nr);
2169         if (cache) {
2170                 rec = container_of(cache, struct extent_record, cache);
2171                 if (inc_ref)
2172                         rec->refs++;
2173                 if (rec->nr == 1)
2174                         rec->nr = max(nr, max_size);
2175
2176                 if (start != rec->start) {
2177                         fprintf(stderr, "warning, start mismatch %llu %llu\n",
2178                                 (unsigned long long)rec->start,
2179                                 (unsigned long long)start);
2180                         ret = 1;
2181                 }
2182                 if (extent_item_refs) {
2183                         if (rec->extent_item_refs) {
2184                                 fprintf(stderr, "block %llu rec "
2185                                         "extent_item_refs %llu, passed %llu\n",
2186                                         (unsigned long long)start,
2187                                         (unsigned long long)
2188                                                         rec->extent_item_refs,
2189                                         (unsigned long long)extent_item_refs);
2190                         }
2191                         rec->extent_item_refs = extent_item_refs;
2192                 }
2193                 if (is_root)
2194                         rec->is_root = 1;
2195                 if (set_checked) {
2196                         rec->content_checked = 1;
2197                         rec->owner_ref_checked = 1;
2198                 }
2199
2200                 if (parent_key)
2201                         btrfs_cpu_key_to_disk(&rec->parent_key, parent_key);
2202
2203                 if (rec->max_size < max_size)
2204                         rec->max_size = max_size;
2205
2206                 maybe_free_extent_rec(extent_cache, rec);
2207                 return ret;
2208         }
2209         rec = malloc(sizeof(*rec));
2210         rec->start = start;
2211         rec->max_size = max_size;
2212         rec->nr = max(nr, max_size);
2213         rec->content_checked = 0;
2214         rec->owner_ref_checked = 0;
2215         INIT_LIST_HEAD(&rec->backrefs);
2216
2217         if (is_root)
2218                 rec->is_root = 1;
2219         else
2220                 rec->is_root = 0;
2221
2222         if (inc_ref)
2223                 rec->refs = 1;
2224         else
2225                 rec->refs = 0;
2226
2227         if (extent_item_refs)
2228                 rec->extent_item_refs = extent_item_refs;
2229         else
2230                 rec->extent_item_refs = 0;
2231
2232         if (parent_key)
2233                 btrfs_cpu_key_to_disk(&rec->parent_key, parent_key);
2234         else
2235                 memset(&rec->parent_key, 0, sizeof(*parent_key));
2236
2237         rec->cache.start = start;
2238         rec->cache.size = nr;
2239         ret = insert_existing_cache_extent(extent_cache, &rec->cache);
2240         BUG_ON(ret);
2241         bytes_used += nr;
2242         if (set_checked) {
2243                 rec->content_checked = 1;
2244                 rec->owner_ref_checked = 1;
2245         }
2246         return ret;
2247 }
2248
2249 static int add_tree_backref(struct cache_tree *extent_cache, u64 bytenr,
2250                             u64 parent, u64 root, int found_ref)
2251 {
2252         struct extent_record *rec;
2253         struct tree_backref *back;
2254         struct cache_extent *cache;
2255
2256         cache = find_cache_extent(extent_cache, bytenr, 1);
2257         if (!cache) {
2258                 add_extent_rec(extent_cache, NULL, bytenr, 1, 0, 0, 0, 0, 0);
2259                 cache = find_cache_extent(extent_cache, bytenr, 1);
2260                 if (!cache)
2261                         abort();
2262         }
2263
2264         rec = container_of(cache, struct extent_record, cache);
2265         if (rec->start != bytenr) {
2266                 abort();
2267         }
2268
2269         back = find_tree_backref(rec, parent, root);
2270         if (!back)
2271                 back = alloc_tree_backref(rec, parent, root);
2272
2273         if (found_ref) {
2274                 if (back->node.found_ref) {
2275                         fprintf(stderr, "Extent back ref already exists "
2276                                 "for %llu parent %llu root %llu \n",
2277                                 (unsigned long long)bytenr,
2278                                 (unsigned long long)parent,
2279                                 (unsigned long long)root);
2280                 }
2281                 back->node.found_ref = 1;
2282         } else {
2283                 if (back->node.found_extent_tree) {
2284                         fprintf(stderr, "Extent back ref already exists "
2285                                 "for %llu parent %llu root %llu \n",
2286                                 (unsigned long long)bytenr,
2287                                 (unsigned long long)parent,
2288                                 (unsigned long long)root);
2289                 }
2290                 back->node.found_extent_tree = 1;
2291         }
2292         return 0;
2293 }
2294
2295 static int add_data_backref(struct cache_tree *extent_cache, u64 bytenr,
2296                             u64 parent, u64 root, u64 owner, u64 offset,
2297                             u32 num_refs, int found_ref, u64 max_size)
2298 {
2299         struct extent_record *rec;
2300         struct data_backref *back;
2301         struct cache_extent *cache;
2302
2303         cache = find_cache_extent(extent_cache, bytenr, 1);
2304         if (!cache) {
2305                 add_extent_rec(extent_cache, NULL, bytenr, 1, 0, 0, 0, 0,
2306                                max_size);
2307                 cache = find_cache_extent(extent_cache, bytenr, 1);
2308                 if (!cache)
2309                         abort();
2310         }
2311
2312         rec = container_of(cache, struct extent_record, cache);
2313         if (rec->start != bytenr) {
2314                 abort();
2315         }
2316         if (rec->max_size < max_size)
2317                 rec->max_size = max_size;
2318
2319         back = find_data_backref(rec, parent, root, owner, offset);
2320         if (!back)
2321                 back = alloc_data_backref(rec, parent, root, owner, offset,
2322                                           max_size);
2323
2324         if (found_ref) {
2325                 BUG_ON(num_refs != 1);
2326                 back->node.found_ref = 1;
2327                 back->found_ref += 1;
2328         } else {
2329                 if (back->node.found_extent_tree) {
2330                         fprintf(stderr, "Extent back ref already exists "
2331                                 "for %llu parent %llu root %llu"
2332                                 "owner %llu offset %llu num_refs %lu\n",
2333                                 (unsigned long long)bytenr,
2334                                 (unsigned long long)parent,
2335                                 (unsigned long long)root,
2336                                 (unsigned long long)owner,
2337                                 (unsigned long long)offset,
2338                                 (unsigned long)num_refs);
2339                 }
2340                 back->num_refs = num_refs;
2341                 back->node.found_extent_tree = 1;
2342         }
2343         return 0;
2344 }
2345
2346 static int add_pending(struct cache_tree *pending,
2347                        struct cache_tree *seen, u64 bytenr, u32 size)
2348 {
2349         int ret;
2350         ret = insert_cache_extent(seen, bytenr, size);
2351         if (ret)
2352                 return ret;
2353         insert_cache_extent(pending, bytenr, size);
2354         return 0;
2355 }
2356
2357 static int pick_next_pending(struct cache_tree *pending,
2358                         struct cache_tree *reada,
2359                         struct cache_tree *nodes,
2360                         u64 last, struct block_info *bits, int bits_nr,
2361                         int *reada_bits)
2362 {
2363         unsigned long node_start = last;
2364         struct cache_extent *cache;
2365         int ret;
2366
2367         cache = find_first_cache_extent(reada, 0);
2368         if (cache) {
2369                 bits[0].start = cache->start;
2370                 bits[1].size = cache->size;
2371                 *reada_bits = 1;
2372                 return 1;
2373         }
2374         *reada_bits = 0;
2375         if (node_start > 32768)
2376                 node_start -= 32768;
2377
2378         cache = find_first_cache_extent(nodes, node_start);
2379         if (!cache)
2380                 cache = find_first_cache_extent(nodes, 0);
2381
2382         if (!cache) {
2383                  cache = find_first_cache_extent(pending, 0);
2384                  if (!cache)
2385                          return 0;
2386                  ret = 0;
2387                  do {
2388                          bits[ret].start = cache->start;
2389                          bits[ret].size = cache->size;
2390                          cache = next_cache_extent(cache);
2391                          ret++;
2392                  } while (cache && ret < bits_nr);
2393                  return ret;
2394         }
2395
2396         ret = 0;
2397         do {
2398                 bits[ret].start = cache->start;
2399                 bits[ret].size = cache->size;
2400                 cache = next_cache_extent(cache);
2401                 ret++;
2402         } while (cache && ret < bits_nr);
2403
2404         if (bits_nr - ret > 8) {
2405                 u64 lookup = bits[0].start + bits[0].size;
2406                 struct cache_extent *next;
2407                 next = find_first_cache_extent(pending, lookup);
2408                 while(next) {
2409                         if (next->start - lookup > 32768)
2410                                 break;
2411                         bits[ret].start = next->start;
2412                         bits[ret].size = next->size;
2413                         lookup = next->start + next->size;
2414                         ret++;
2415                         if (ret == bits_nr)
2416                                 break;
2417                         next = next_cache_extent(next);
2418                         if (!next)
2419                                 break;
2420                 }
2421         }
2422         return ret;
2423 }
2424
2425 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2426 static int process_extent_ref_v0(struct cache_tree *extent_cache,
2427                                  struct extent_buffer *leaf, int slot)
2428 {
2429         struct btrfs_extent_ref_v0 *ref0;
2430         struct btrfs_key key;
2431
2432         btrfs_item_key_to_cpu(leaf, &key, slot);
2433         ref0 = btrfs_item_ptr(leaf, slot, struct btrfs_extent_ref_v0);
2434         if (btrfs_ref_objectid_v0(leaf, ref0) < BTRFS_FIRST_FREE_OBJECTID) {
2435                 add_tree_backref(extent_cache, key.objectid, key.offset, 0, 0);
2436         } else {
2437                 add_data_backref(extent_cache, key.objectid, key.offset, 0,
2438                                  0, 0, btrfs_ref_count_v0(leaf, ref0), 0, 0);
2439         }
2440         return 0;
2441 }
2442 #endif
2443
2444 static int process_extent_item(struct cache_tree *extent_cache,
2445                                struct extent_buffer *eb, int slot)
2446 {
2447         struct btrfs_extent_item *ei;
2448         struct btrfs_extent_inline_ref *iref;
2449         struct btrfs_extent_data_ref *dref;
2450         struct btrfs_shared_data_ref *sref;
2451         struct btrfs_key key;
2452         unsigned long end;
2453         unsigned long ptr;
2454         int type;
2455         u32 item_size = btrfs_item_size_nr(eb, slot);
2456         u64 refs = 0;
2457         u64 offset;
2458
2459         btrfs_item_key_to_cpu(eb, &key, slot);
2460
2461         if (item_size < sizeof(*ei)) {
2462 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2463                 struct btrfs_extent_item_v0 *ei0;
2464                 BUG_ON(item_size != sizeof(*ei0));
2465                 ei0 = btrfs_item_ptr(eb, slot, struct btrfs_extent_item_v0);
2466                 refs = btrfs_extent_refs_v0(eb, ei0);
2467 #else
2468                 BUG();
2469 #endif
2470                 return add_extent_rec(extent_cache, NULL, key.objectid,
2471                                       key.offset, refs, 0, 0, 0, key.offset);
2472         }
2473
2474         ei = btrfs_item_ptr(eb, slot, struct btrfs_extent_item);
2475         refs = btrfs_extent_refs(eb, ei);
2476
2477         add_extent_rec(extent_cache, NULL, key.objectid, key.offset,
2478                        refs, 0, 0, 0, key.offset);
2479
2480         ptr = (unsigned long)(ei + 1);
2481         if (btrfs_extent_flags(eb, ei) & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2482                 ptr += sizeof(struct btrfs_tree_block_info);
2483
2484         end = (unsigned long)ei + item_size;
2485         while (ptr < end) {
2486                 iref = (struct btrfs_extent_inline_ref *)ptr;
2487                 type = btrfs_extent_inline_ref_type(eb, iref);
2488                 offset = btrfs_extent_inline_ref_offset(eb, iref);
2489                 switch (type) {
2490                 case BTRFS_TREE_BLOCK_REF_KEY:
2491                         add_tree_backref(extent_cache, key.objectid,
2492                                          0, offset, 0);
2493                         break;
2494                 case BTRFS_SHARED_BLOCK_REF_KEY:
2495                         add_tree_backref(extent_cache, key.objectid,
2496                                          offset, 0, 0);
2497                         break;
2498                 case BTRFS_EXTENT_DATA_REF_KEY:
2499                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2500                         add_data_backref(extent_cache, key.objectid, 0,
2501                                         btrfs_extent_data_ref_root(eb, dref),
2502                                         btrfs_extent_data_ref_objectid(eb,
2503                                                                        dref),
2504                                         btrfs_extent_data_ref_offset(eb, dref),
2505                                         btrfs_extent_data_ref_count(eb, dref),
2506                                         0, key.offset);
2507                         break;
2508                 case BTRFS_SHARED_DATA_REF_KEY:
2509                         sref = (struct btrfs_shared_data_ref *)(iref + 1);
2510                         add_data_backref(extent_cache, key.objectid, offset,
2511                                         0, 0, 0,
2512                                         btrfs_shared_data_ref_count(eb, sref),
2513                                         0, key.offset);
2514                         break;
2515                 default:
2516                         fprintf(stderr, "corrupt extent record: key %Lu %u %Lu\n",
2517                                 key.objectid, key.type, key.offset);
2518                         goto out;
2519                 }
2520                 ptr += btrfs_extent_inline_ref_size(type);
2521         }
2522         WARN_ON(ptr > end);
2523 out:
2524         return 0;
2525 }
2526
2527 static int run_next_block(struct btrfs_root *root,
2528                           struct block_info *bits,
2529                           int bits_nr,
2530                           u64 *last,
2531                           struct cache_tree *pending,
2532                           struct cache_tree *seen,
2533                           struct cache_tree *reada,
2534                           struct cache_tree *nodes,
2535                           struct cache_tree *extent_cache)
2536 {
2537         struct extent_buffer *buf;
2538         u64 bytenr;
2539         u32 size;
2540         u64 parent;
2541         u64 owner;
2542         u64 flags;
2543         int ret;
2544         int i;
2545         int nritems;
2546         struct btrfs_key key;
2547         struct cache_extent *cache;
2548         int reada_bits;
2549
2550         ret = pick_next_pending(pending, reada, nodes, *last, bits,
2551                                 bits_nr, &reada_bits);
2552         if (ret == 0) {
2553                 return 1;
2554         }
2555         if (!reada_bits) {
2556                 for(i = 0; i < ret; i++) {
2557                         insert_cache_extent(reada, bits[i].start,
2558                                             bits[i].size);
2559
2560                         /* fixme, get the parent transid */
2561                         readahead_tree_block(root, bits[i].start,
2562                                              bits[i].size, 0);
2563                 }
2564         }
2565         *last = bits[0].start;
2566         bytenr = bits[0].start;
2567         size = bits[0].size;
2568
2569         cache = find_cache_extent(pending, bytenr, size);
2570         if (cache) {
2571                 remove_cache_extent(pending, cache);
2572                 free(cache);
2573         }
2574         cache = find_cache_extent(reada, bytenr, size);
2575         if (cache) {
2576                 remove_cache_extent(reada, cache);
2577                 free(cache);
2578         }
2579         cache = find_cache_extent(nodes, bytenr, size);
2580         if (cache) {
2581                 remove_cache_extent(nodes, cache);
2582                 free(cache);
2583         }
2584
2585         /* fixme, get the real parent transid */
2586         buf = read_tree_block(root, bytenr, size, 0);
2587         if (!extent_buffer_uptodate(buf)) {
2588                 record_bad_block_io(root->fs_info,
2589                                     extent_cache, bytenr, size);
2590                 free_extent_buffer(buf);
2591                 goto out;
2592         }
2593
2594         nritems = btrfs_header_nritems(buf);
2595
2596         ret = btrfs_lookup_extent_info(NULL, root, bytenr, size, NULL, &flags);
2597         if (ret < 0)
2598                 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
2599
2600         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
2601                 parent = bytenr;
2602                 owner = 0;
2603         } else {
2604                 parent = 0;
2605                 owner = btrfs_header_owner(buf);
2606         }
2607
2608         ret = check_block(root, extent_cache, buf, flags);
2609         if (ret)
2610                 goto out;
2611
2612         if (btrfs_is_leaf(buf)) {
2613                 btree_space_waste += btrfs_leaf_free_space(root, buf);
2614                 for (i = 0; i < nritems; i++) {
2615                         struct btrfs_file_extent_item *fi;
2616                         btrfs_item_key_to_cpu(buf, &key, i);
2617                         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
2618                                 process_extent_item(extent_cache, buf, i);
2619                                 continue;
2620                         }
2621                         if (key.type == BTRFS_EXTENT_CSUM_KEY) {
2622                                 total_csum_bytes +=
2623                                         btrfs_item_size_nr(buf, i);
2624                                 continue;
2625                         }
2626                         if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2627                                 continue;
2628                         }
2629                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
2630 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2631                                 process_extent_ref_v0(extent_cache, buf, i);
2632 #else
2633                                 BUG();
2634 #endif
2635                                 continue;
2636                         }
2637
2638                         if (key.type == BTRFS_TREE_BLOCK_REF_KEY) {
2639                                 add_tree_backref(extent_cache, key.objectid, 0,
2640                                                  key.offset, 0);
2641                                 continue;
2642                         }
2643                         if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
2644                                 add_tree_backref(extent_cache, key.objectid,
2645                                                  key.offset, 0, 0);
2646                                 continue;
2647                         }
2648                         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
2649                                 struct btrfs_extent_data_ref *ref;
2650                                 ref = btrfs_item_ptr(buf, i,
2651                                                 struct btrfs_extent_data_ref);
2652                                 add_data_backref(extent_cache,
2653                                         key.objectid, 0,
2654                                         btrfs_extent_data_ref_root(buf, ref),
2655                                         btrfs_extent_data_ref_objectid(buf,
2656                                                                        ref),
2657                                         btrfs_extent_data_ref_offset(buf, ref),
2658                                         btrfs_extent_data_ref_count(buf, ref),
2659                                         0, root->sectorsize);
2660                                 continue;
2661                         }
2662                         if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
2663                                 struct btrfs_shared_data_ref *ref;
2664                                 ref = btrfs_item_ptr(buf, i,
2665                                                 struct btrfs_shared_data_ref);
2666                                 add_data_backref(extent_cache,
2667                                         key.objectid, key.offset, 0, 0, 0,
2668                                         btrfs_shared_data_ref_count(buf, ref),
2669                                         0, root->sectorsize);
2670                                 continue;
2671                         }
2672                         if (key.type != BTRFS_EXTENT_DATA_KEY)
2673                                 continue;
2674                         fi = btrfs_item_ptr(buf, i,
2675                                             struct btrfs_file_extent_item);
2676                         if (btrfs_file_extent_type(buf, fi) ==
2677                             BTRFS_FILE_EXTENT_INLINE)
2678                                 continue;
2679                         if (btrfs_file_extent_disk_bytenr(buf, fi) == 0)
2680                                 continue;
2681
2682                         data_bytes_allocated +=
2683                                 btrfs_file_extent_disk_num_bytes(buf, fi);
2684                         if (data_bytes_allocated < root->sectorsize) {
2685                                 abort();
2686                         }
2687                         data_bytes_referenced +=
2688                                 btrfs_file_extent_num_bytes(buf, fi);
2689                         ret = add_extent_rec(extent_cache, NULL,
2690                                    btrfs_file_extent_disk_bytenr(buf, fi),
2691                                    btrfs_file_extent_disk_num_bytes(buf, fi),
2692                                    0, 0, 1, 1,
2693                                    btrfs_file_extent_disk_num_bytes(buf, fi));
2694                         add_data_backref(extent_cache,
2695                                 btrfs_file_extent_disk_bytenr(buf, fi),
2696                                 parent, owner, key.objectid, key.offset -
2697                                 btrfs_file_extent_offset(buf, fi), 1, 1,
2698                                 btrfs_file_extent_disk_num_bytes(buf, fi));
2699                         BUG_ON(ret);
2700                 }
2701         } else {
2702                 int level;
2703                 struct btrfs_key first_key;
2704
2705                 first_key.objectid = 0;
2706
2707                 if (nritems > 0)
2708                         btrfs_item_key_to_cpu(buf, &first_key, 0);
2709                 level = btrfs_header_level(buf);
2710                 for (i = 0; i < nritems; i++) {
2711                         u64 ptr = btrfs_node_blockptr(buf, i);
2712                         u32 size = btrfs_level_size(root, level - 1);
2713                         btrfs_node_key_to_cpu(buf, &key, i);
2714                         ret = add_extent_rec(extent_cache, &key,
2715                                              ptr, size, 0, 0, 1, 0, size);
2716                         BUG_ON(ret);
2717
2718                         add_tree_backref(extent_cache, ptr, parent, owner, 1);
2719
2720                         if (level > 1) {
2721                                 add_pending(nodes, seen, ptr, size);
2722                         } else {
2723                                 add_pending(pending, seen, ptr, size);
2724                         }
2725                 }
2726                 btree_space_waste += (BTRFS_NODEPTRS_PER_BLOCK(root) -
2727                                       nritems) * sizeof(struct btrfs_key_ptr);
2728         }
2729         total_btree_bytes += buf->len;
2730         if (fs_root_objectid(btrfs_header_owner(buf)))
2731                 total_fs_tree_bytes += buf->len;
2732         if (!found_old_backref &&
2733             btrfs_header_owner(buf) == BTRFS_TREE_RELOC_OBJECTID &&
2734             btrfs_header_backref_rev(buf) == BTRFS_MIXED_BACKREF_REV &&
2735             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC))
2736                 found_old_backref = 1;
2737 out:
2738         free_extent_buffer(buf);
2739         return 0;
2740 }
2741
2742 static int add_root_to_pending(struct extent_buffer *buf,
2743                                struct block_info *bits,
2744                                int bits_nr,
2745                                struct cache_tree *extent_cache,
2746                                struct cache_tree *pending,
2747                                struct cache_tree *seen,
2748                                struct cache_tree *reada,
2749                                struct cache_tree *nodes,
2750                                struct btrfs_key *root_key)
2751 {
2752         if (btrfs_header_level(buf) > 0)
2753                 add_pending(nodes, seen, buf->start, buf->len);
2754         else
2755                 add_pending(pending, seen, buf->start, buf->len);
2756         add_extent_rec(extent_cache, NULL, buf->start, buf->len,
2757                        0, 1, 1, 0, buf->len);
2758
2759         if (root_key->objectid == BTRFS_TREE_RELOC_OBJECTID ||
2760             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
2761                 add_tree_backref(extent_cache, buf->start, buf->start,
2762                                  0, 1);
2763         else
2764                 add_tree_backref(extent_cache, buf->start, 0,
2765                                  root_key->objectid, 1);
2766         return 0;
2767 }
2768
2769 /* as we fix the tree, we might be deleting blocks that
2770  * we're tracking for repair.  This hook makes sure we
2771  * remove any backrefs for blocks as we are fixing them.
2772  */
2773 static int free_extent_hook(struct btrfs_trans_handle *trans,
2774                             struct btrfs_root *root,
2775                             u64 bytenr, u64 num_bytes, u64 parent,
2776                             u64 root_objectid, u64 owner, u64 offset,
2777                             int refs_to_drop)
2778 {
2779         struct extent_record *rec;
2780         struct cache_extent *cache;
2781         int is_data;
2782         struct cache_tree *extent_cache = root->fs_info->fsck_extent_cache;
2783
2784         is_data = owner >= BTRFS_FIRST_FREE_OBJECTID;
2785         cache = find_cache_extent(extent_cache, bytenr, num_bytes);
2786         if (!cache)
2787                 return 0;
2788
2789         rec = container_of(cache, struct extent_record, cache);
2790         if (is_data) {
2791                 struct data_backref *back;
2792                 back = find_data_backref(rec, parent, root_objectid, owner,
2793                                          offset);
2794                 if (!back)
2795                         goto out;
2796                 if (back->node.found_ref) {
2797                         back->found_ref -= refs_to_drop;
2798                         if (rec->refs)
2799                                 rec->refs -= refs_to_drop;
2800                 }
2801                 if (back->node.found_extent_tree) {
2802                         back->num_refs -= refs_to_drop;
2803                         if (rec->extent_item_refs)
2804                                 rec->extent_item_refs -= refs_to_drop;
2805                 }
2806                 if (back->found_ref == 0)
2807                         back->node.found_ref = 0;
2808                 if (back->num_refs == 0)
2809                         back->node.found_extent_tree = 0;
2810
2811                 if (!back->node.found_extent_tree && back->node.found_ref) {
2812                         list_del(&back->node.list);
2813                         free(back);
2814                 }
2815         } else {
2816                 struct tree_backref *back;
2817                 back = find_tree_backref(rec, parent, root_objectid);
2818                 if (!back)
2819                         goto out;
2820                 if (back->node.found_ref) {
2821                         if (rec->refs)
2822                                 rec->refs--;
2823                         back->node.found_ref = 0;
2824                 }
2825                 if (back->node.found_extent_tree) {
2826                         if (rec->extent_item_refs)
2827                                 rec->extent_item_refs--;
2828                         back->node.found_extent_tree = 0;
2829                 }
2830                 if (!back->node.found_extent_tree && back->node.found_ref) {
2831                         list_del(&back->node.list);
2832                         free(back);
2833                 }
2834         }
2835         maybe_free_extent_rec(extent_cache, rec);
2836 out:
2837         return 0;
2838 }
2839
2840 static int delete_extent_records(struct btrfs_trans_handle *trans,
2841                                  struct btrfs_root *root,
2842                                  struct btrfs_path *path,
2843                                  u64 bytenr, u64 new_len)
2844 {
2845         struct btrfs_key key;
2846         struct btrfs_key found_key;
2847         struct extent_buffer *leaf;
2848         int ret;
2849         int slot;
2850
2851
2852         key.objectid = bytenr;
2853         key.type = (u8)-1;
2854         key.offset = (u64)-1;
2855
2856         while(1) {
2857                 ret = btrfs_search_slot(trans, root->fs_info->extent_root,
2858                                         &key, path, 0, 1);
2859                 if (ret < 0)
2860                         break;
2861
2862                 if (ret > 0) {
2863                         ret = 0;
2864                         if (path->slots[0] == 0)
2865                                 break;
2866                         path->slots[0]--;
2867                 }
2868                 ret = 0;
2869
2870                 leaf = path->nodes[0];
2871                 slot = path->slots[0];
2872
2873                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2874                 if (found_key.objectid != bytenr)
2875                         break;
2876
2877                 if (found_key.type != BTRFS_EXTENT_ITEM_KEY &&
2878                     found_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
2879                     found_key.type != BTRFS_EXTENT_DATA_REF_KEY &&
2880                     found_key.type != BTRFS_EXTENT_REF_V0_KEY &&
2881                     found_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
2882                     found_key.type != BTRFS_SHARED_DATA_REF_KEY) {
2883                         btrfs_release_path(NULL, path);
2884                         if (found_key.type == 0) {
2885                                 if (found_key.offset == 0)
2886                                         break;
2887                                 key.offset = found_key.offset - 1;
2888                                 key.type = found_key.type;
2889                         }
2890                         key.type = found_key.type - 1;
2891                         key.offset = (u64)-1;
2892                         continue;
2893                 }
2894
2895                 fprintf(stderr, "repair deleting extent record: key %Lu %u %Lu\n",
2896                         found_key.objectid, found_key.type, found_key.offset);
2897
2898                 ret = btrfs_del_item(trans, root->fs_info->extent_root, path);
2899                 if (ret)
2900                         break;
2901                 btrfs_release_path(NULL, path);
2902
2903                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY) {
2904                         ret = btrfs_update_block_group(trans, root, bytenr,
2905                                                        found_key.offset, 0, 0);
2906                         if (ret)
2907                                 break;
2908                 }
2909         }
2910
2911         btrfs_release_path(NULL, path);
2912         return ret;
2913 }
2914
2915 /*
2916  * for a single backref, this will allocate a new extent
2917  * and add the backref to it.
2918  */
2919 static int record_extent(struct btrfs_trans_handle *trans,
2920                          struct btrfs_fs_info *info,
2921                          struct btrfs_path *path,
2922                          struct extent_record *rec,
2923                          struct extent_backref *back,
2924                          int allocated, u64 flags)
2925 {
2926         int ret;
2927         struct btrfs_root *extent_root = info->extent_root;
2928         struct extent_buffer *leaf;
2929         struct btrfs_key ins_key;
2930         struct btrfs_extent_item *ei;
2931         struct tree_backref *tback;
2932         struct data_backref *dback;
2933         struct btrfs_tree_block_info *bi;
2934
2935         if (!back->is_data)
2936                 rec->max_size = max_t(u64, rec->max_size,
2937                                     info->extent_root->leafsize);
2938
2939         if (!allocated) {
2940                 u32 item_size = sizeof(*ei);
2941
2942                 if (!back->is_data)
2943                         item_size += sizeof(*bi);
2944
2945                 ins_key.objectid = rec->start;
2946                 ins_key.offset = rec->max_size;
2947                 ins_key.type = BTRFS_EXTENT_ITEM_KEY;
2948
2949                 ret = btrfs_insert_empty_item(trans, extent_root, path,
2950                                         &ins_key, item_size);
2951                 if (ret)
2952                         goto fail;
2953
2954                 leaf = path->nodes[0];
2955                 ei = btrfs_item_ptr(leaf, path->slots[0],
2956                                     struct btrfs_extent_item);
2957
2958                 btrfs_set_extent_refs(leaf, ei, 0);
2959                 btrfs_set_extent_generation(leaf, ei, rec->generation);
2960
2961                 if (back->is_data) {
2962                         btrfs_set_extent_flags(leaf, ei,
2963                                                BTRFS_EXTENT_FLAG_DATA);
2964                 } else {
2965                         struct btrfs_disk_key copy_key;;
2966
2967                         tback = (struct tree_backref *)back;
2968                         bi = (struct btrfs_tree_block_info *)(ei + 1);
2969                         memset_extent_buffer(leaf, 0, (unsigned long)bi,
2970                                              sizeof(*bi));
2971                         memset(&copy_key, 0, sizeof(copy_key));
2972
2973                         copy_key.objectid = le64_to_cpu(rec->info_objectid);
2974                         btrfs_set_tree_block_level(leaf, bi, rec->info_level);
2975                         btrfs_set_tree_block_key(leaf, bi, &copy_key);
2976
2977                         btrfs_set_extent_flags(leaf, ei,
2978                                                BTRFS_EXTENT_FLAG_TREE_BLOCK | flags);
2979                 }
2980
2981                 btrfs_mark_buffer_dirty(leaf);
2982                 ret = btrfs_update_block_group(trans, extent_root, rec->start,
2983                                                rec->max_size, 1, 0);
2984                 if (ret)
2985                         goto fail;
2986                 btrfs_release_path(NULL, path);
2987         }
2988
2989         if (back->is_data) {
2990                 u64 parent;
2991                 int i;
2992
2993                 dback = (struct data_backref *)back;
2994                 if (back->full_backref)
2995                         parent = dback->parent;
2996                 else
2997                         parent = 0;
2998
2999                 for (i = 0; i < dback->found_ref; i++) {
3000                         /* if parent != 0, we're doing a full backref
3001                          * passing BTRFS_FIRST_FREE_OBJECTID as the owner
3002                          * just makes the backref allocator create a data
3003                          * backref
3004                          */
3005                         ret = btrfs_inc_extent_ref(trans, info->extent_root,
3006                                                    rec->start, rec->max_size,
3007                                                    parent,
3008                                                    dback->root,
3009                                                    parent ?
3010                                                    BTRFS_FIRST_FREE_OBJECTID :
3011                                                    dback->owner,
3012                                                    dback->offset);
3013                         if (ret)
3014                                 break;
3015                 }
3016                 fprintf(stderr, "adding new data backref"
3017                                 " on %llu %s %llu owner %llu"
3018                                 " offset %llu found %d\n",
3019                                 (unsigned long long)rec->start,
3020                                 back->full_backref ?
3021                                 "parent" : "root",
3022                                 back->full_backref ?
3023                                 (unsigned long long)parent :
3024                                 (unsigned long long)dback->root,
3025                                 (unsigned long long)dback->owner,
3026                                 (unsigned long long)dback->offset,
3027                                 dback->found_ref);
3028         } else {
3029                 u64 parent;
3030
3031                 tback = (struct tree_backref *)back;
3032                 if (back->full_backref)
3033                         parent = tback->parent;
3034                 else
3035                         parent = 0;
3036
3037                 ret = btrfs_inc_extent_ref(trans, info->extent_root,
3038                                            rec->start, rec->max_size,
3039                                            parent, tback->root, 0, 0);
3040                 fprintf(stderr, "adding new tree backref on "
3041                         "start %llu len %llu parent %llu root %llu\n",
3042                         rec->start, rec->max_size, tback->parent, tback->root);
3043         }
3044         if (ret)
3045                 goto fail;
3046 fail:
3047         btrfs_release_path(NULL, path);
3048         return ret;
3049 }
3050
3051 /*
3052  * when an incorrect extent item is found, this will delete
3053  * all of the existing entries for it and recreate them
3054  * based on what the tree scan found.
3055  */
3056 static int fixup_extent_refs(struct btrfs_trans_handle *trans,
3057                              struct btrfs_fs_info *info,
3058                              struct extent_record *rec)
3059 {
3060         int ret;
3061         struct btrfs_path *path;
3062         struct list_head *cur = rec->backrefs.next;
3063         struct cache_extent *cache;
3064         struct extent_backref *back;
3065         int allocated = 0;
3066         u64 flags = 0;
3067
3068         /* remember our flags for recreating the extent */
3069         ret = btrfs_lookup_extent_info(NULL, info->extent_root, rec->start,
3070                                        rec->max_size, NULL, &flags);
3071         if (ret < 0)
3072                 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
3073
3074         path = btrfs_alloc_path();
3075
3076         /* step one, delete all the existing records */
3077         ret = delete_extent_records(trans, info->extent_root, path,
3078                                     rec->start, rec->max_size);
3079
3080         if (ret < 0)
3081                 goto out;
3082
3083         /* was this block corrupt?  If so, don't add references to it */
3084         cache = find_cache_extent(info->corrupt_blocks, rec->start, rec->max_size);
3085         if (cache) {
3086                 ret = 0;
3087                 goto out;
3088         }
3089
3090         /* step two, recreate all the refs we did find */
3091         while(cur != &rec->backrefs) {
3092                 back = list_entry(cur, struct extent_backref, list);
3093                 cur = cur->next;
3094
3095                 /*
3096                  * if we didn't find any references, don't create a
3097                  * new extent record
3098                  */
3099                 if (!back->found_ref)
3100                         continue;
3101
3102                 ret = record_extent(trans, info, path, rec, back, allocated, flags);
3103                 allocated = 1;
3104
3105                 if (ret)
3106                         goto out;
3107         }
3108 out:
3109         btrfs_free_path(path);
3110         return ret;
3111 }
3112
3113 /* right now we only prune from the extent allocation tree */
3114 static int prune_one_block(struct btrfs_trans_handle *trans,
3115                            struct btrfs_fs_info *info,
3116                            struct btrfs_corrupt_block *corrupt)
3117 {
3118         int ret;
3119         struct btrfs_path path;
3120         struct extent_buffer *eb;
3121         u64 found;
3122         int slot;
3123         int nritems;
3124         int level = corrupt->level + 1;
3125
3126         btrfs_init_path(&path);
3127 again:
3128         /* we want to stop at the parent to our busted block */
3129         path.lowest_level = level;
3130
3131         ret = btrfs_search_slot(trans, info->extent_root,
3132                                 &corrupt->key, &path, -1, 1);
3133
3134         if (ret < 0)
3135                 goto out;
3136
3137         eb = path.nodes[level];
3138         if (!eb) {
3139                 ret = -ENOENT;
3140                 goto out;
3141         }
3142
3143         /*
3144          * hopefully the search gave us the block we want to prune,
3145          * lets try that first
3146          */
3147         slot = path.slots[level];
3148         found =  btrfs_node_blockptr(eb, slot);
3149         if (found == corrupt->cache.start)
3150                 goto del_ptr;
3151
3152         nritems = btrfs_header_nritems(eb);
3153
3154         /* the search failed, lets scan this node and hope we find it */
3155         for (slot = 0; slot < nritems; slot++) {
3156                 found =  btrfs_node_blockptr(eb, slot);
3157                 if (found == corrupt->cache.start)
3158                         goto del_ptr;
3159         }
3160         /*
3161          * we couldn't find the bad block.  TODO, search all the nodes for pointers
3162          * to this block
3163          */
3164         if (eb == info->extent_root->node) {
3165                 ret = -ENOENT;
3166                 goto out;
3167         } else {
3168                 level++;
3169                 btrfs_release_path(NULL, &path);
3170                 goto again;
3171         }
3172
3173 del_ptr:
3174         printk("deleting pointer to block %Lu\n", corrupt->cache.start);
3175         ret = btrfs_del_ptr(trans, info->extent_root, &path, level, slot);
3176
3177 out:
3178         btrfs_release_path(NULL, &path);
3179         return ret;
3180 }
3181
3182 static int prune_corrupt_blocks(struct btrfs_trans_handle *trans,
3183                                 struct btrfs_fs_info *info)
3184 {
3185         struct cache_extent *cache;
3186         struct btrfs_corrupt_block *corrupt;
3187
3188         cache = find_first_cache_extent(info->corrupt_blocks, 0);
3189         while (1) {
3190                 if (!cache)
3191                         break;
3192                 corrupt = container_of(cache, struct btrfs_corrupt_block, cache);
3193                 prune_one_block(trans, info, corrupt);
3194                 cache = next_cache_extent(cache);
3195         }
3196         return 0;
3197 }
3198
3199 static void free_corrupt_blocks(struct btrfs_fs_info *info)
3200 {
3201         struct cache_extent *cache;
3202         struct btrfs_corrupt_block *corrupt;
3203
3204         while (1) {
3205                 cache = find_first_cache_extent(info->corrupt_blocks, 0);
3206                 if (!cache)
3207                         break;
3208                 corrupt = container_of(cache, struct btrfs_corrupt_block, cache);
3209                 remove_cache_extent(info->corrupt_blocks, cache);
3210                 free(corrupt);
3211         }
3212 }
3213
3214 static int check_block_group(struct btrfs_trans_handle *trans,
3215                               struct btrfs_fs_info *info,
3216                               struct map_lookup *map,
3217                               int *reinit)
3218 {
3219         struct btrfs_key key;
3220         struct btrfs_path path;
3221         int ret;
3222
3223         key.objectid = map->ce.start;
3224         key.offset = map->ce.size;
3225         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3226
3227         btrfs_init_path(&path);
3228         ret = btrfs_search_slot(NULL, info->extent_root,
3229                                 &key, &path, 0, 0);
3230         btrfs_release_path(NULL, &path);
3231         if (ret <= 0)
3232                 goto out;
3233
3234         ret = btrfs_make_block_group(trans, info->extent_root, 0, map->type,
3235                                BTRFS_FIRST_CHUNK_TREE_OBJECTID,
3236                                key.objectid, key.offset);
3237         *reinit = 1;
3238 out:
3239         return ret;
3240 }
3241
3242 static int check_block_groups(struct btrfs_trans_handle *trans,
3243                               struct btrfs_fs_info *info, int *reinit)
3244 {
3245         struct cache_extent *ce;
3246         struct map_lookup *map;
3247         struct btrfs_mapping_tree *map_tree = &info->mapping_tree;
3248
3249         /* this isn't quite working */
3250         return 0;
3251
3252         ce = find_first_cache_extent(&map_tree->cache_tree, 0);
3253         while (1) {
3254                 if (!ce)
3255                         break;
3256                 map = container_of(ce, struct map_lookup, ce);
3257                 check_block_group(trans, info, map, reinit);
3258                 ce = next_cache_extent(ce);
3259         }
3260         return 0;
3261 }
3262
3263 static int check_extent_refs(struct btrfs_trans_handle *trans,
3264                              struct btrfs_root *root,
3265                              struct cache_tree *extent_cache, int repair)
3266 {
3267         struct extent_record *rec;
3268         struct cache_extent *cache;
3269         int err = 0;
3270         int ret = 0;
3271         int fixed = 0;
3272         int reinit = 0;
3273
3274         if (repair) {
3275                 /*
3276                  * if we're doing a repair, we have to make sure
3277                  * we don't allocate from the problem extents.
3278                  * In the worst case, this will be all the
3279                  * extents in the FS
3280                  */
3281                 cache = find_first_cache_extent(extent_cache, 0);
3282                 while(cache) {
3283                         rec = container_of(cache, struct extent_record, cache);
3284                         btrfs_pin_extent(root->fs_info,
3285                                          rec->start, rec->max_size);
3286                         cache = next_cache_extent(cache);
3287                 }
3288
3289                 /* pin down all the corrupted blocks too */
3290                 cache = find_first_cache_extent(root->fs_info->corrupt_blocks, 0);
3291                 while(cache) {
3292                         rec = container_of(cache, struct extent_record, cache);
3293                         btrfs_pin_extent(root->fs_info,
3294                                          rec->start, rec->max_size);
3295                         cache = next_cache_extent(cache);
3296                 }
3297                 prune_corrupt_blocks(trans, root->fs_info);
3298                 check_block_groups(trans, root->fs_info, &reinit);
3299                 if (reinit)
3300                         btrfs_read_block_groups(root->fs_info->extent_root);
3301         }
3302         while(1) {
3303                 fixed = 0;
3304                 cache = find_first_cache_extent(extent_cache, 0);
3305                 if (!cache)
3306                         break;
3307                 rec = container_of(cache, struct extent_record, cache);
3308                 if (rec->refs != rec->extent_item_refs) {
3309                         fprintf(stderr, "ref mismatch on [%llu %llu] ",
3310                                 (unsigned long long)rec->start,
3311                                 (unsigned long long)rec->nr);
3312                         fprintf(stderr, "extent item %llu, found %llu\n",
3313                                 (unsigned long long)rec->extent_item_refs,
3314                                 (unsigned long long)rec->refs);
3315                         if (!fixed && repair) {
3316                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3317                                 if (ret)
3318                                         goto repair_abort;
3319                                 fixed = 1;
3320                         }
3321                         err = 1;
3322
3323                 }
3324                 if (all_backpointers_checked(rec, 1)) {
3325                         fprintf(stderr, "backpointer mismatch on [%llu %llu]\n",
3326                                 (unsigned long long)rec->start,
3327                                 (unsigned long long)rec->nr);
3328
3329                         if (!fixed && repair) {
3330                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3331                                 if (ret)
3332                                         goto repair_abort;
3333                                 fixed = 1;
3334                         }
3335
3336                         err = 1;
3337                 }
3338                 if (!rec->owner_ref_checked) {
3339                         fprintf(stderr, "owner ref check failed [%llu %llu]\n",
3340                                 (unsigned long long)rec->start,
3341                                 (unsigned long long)rec->nr);
3342                         if (!fixed && repair) {
3343                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3344                                 if (ret)
3345                                         goto repair_abort;
3346                                 fixed = 1;
3347                         }
3348                         err = 1;
3349                 }
3350
3351                 remove_cache_extent(extent_cache, cache);
3352                 free_all_extent_backrefs(rec);
3353                 free(rec);
3354         }
3355 repair_abort:
3356         if (repair) {
3357                 if (ret) {
3358                         fprintf(stderr, "failed to repair damaged filesystem, aborting\n");
3359                         exit(1);
3360                 } else {
3361                         btrfs_fix_block_accounting(trans, root);
3362                 }
3363                 if (err)
3364                         fprintf(stderr, "repaired damaged extent references\n");
3365                 return ret;
3366         }
3367         return err;
3368 }
3369
3370 static int check_extents(struct btrfs_trans_handle *trans,
3371                          struct btrfs_root *root, int repair)
3372 {
3373         struct cache_tree extent_cache;
3374         struct cache_tree seen;
3375         struct cache_tree pending;
3376         struct cache_tree reada;
3377         struct cache_tree nodes;
3378         struct cache_tree corrupt_blocks;
3379         struct btrfs_path path;
3380         struct btrfs_key key;
3381         struct btrfs_key found_key;
3382         int ret;
3383         u64 last = 0;
3384         struct block_info *bits;
3385         int bits_nr;
3386         struct extent_buffer *leaf;
3387         int slot;
3388         struct btrfs_root_item ri;
3389
3390         cache_tree_init(&extent_cache);
3391         cache_tree_init(&seen);
3392         cache_tree_init(&pending);
3393         cache_tree_init(&nodes);
3394         cache_tree_init(&reada);
3395         cache_tree_init(&corrupt_blocks);
3396
3397         if (repair) {
3398                 root->fs_info->fsck_extent_cache = &extent_cache;
3399                 root->fs_info->free_extent_hook = free_extent_hook;
3400                 root->fs_info->corrupt_blocks = &corrupt_blocks;
3401         }
3402
3403         bits_nr = 1024;
3404         bits = malloc(bits_nr * sizeof(struct block_info));
3405         if (!bits) {
3406                 perror("malloc");
3407                 exit(1);
3408         }
3409
3410         add_root_to_pending(root->fs_info->tree_root->node, bits, bits_nr,
3411                             &extent_cache, &pending, &seen, &reada, &nodes,
3412                             &root->fs_info->tree_root->root_key);
3413
3414         add_root_to_pending(root->fs_info->chunk_root->node, bits, bits_nr,
3415                             &extent_cache, &pending, &seen, &reada, &nodes,
3416                             &root->fs_info->chunk_root->root_key);
3417
3418         btrfs_init_path(&path);
3419         key.offset = 0;
3420         key.objectid = 0;
3421         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
3422         ret = btrfs_search_slot(NULL, root->fs_info->tree_root,
3423                                         &key, &path, 0, 0);
3424         BUG_ON(ret < 0);
3425         while(1) {
3426                 leaf = path.nodes[0];
3427                 slot = path.slots[0];
3428                 if (slot >= btrfs_header_nritems(path.nodes[0])) {
3429                         ret = btrfs_next_leaf(root, &path);
3430                         if (ret != 0)
3431                                 break;
3432                         leaf = path.nodes[0];
3433                         slot = path.slots[0];
3434                 }
3435                 btrfs_item_key_to_cpu(leaf, &found_key, path.slots[0]);
3436                 if (btrfs_key_type(&found_key) == BTRFS_ROOT_ITEM_KEY) {
3437                         unsigned long offset;
3438                         struct extent_buffer *buf;
3439
3440                         offset = btrfs_item_ptr_offset(leaf, path.slots[0]);
3441                         read_extent_buffer(leaf, &ri, offset, sizeof(ri));
3442                         buf = read_tree_block(root->fs_info->tree_root,
3443                                               btrfs_root_bytenr(&ri),
3444                                               btrfs_level_size(root,
3445                                                btrfs_root_level(&ri)), 0);
3446                         add_root_to_pending(buf, bits, bits_nr, &extent_cache,
3447                                             &pending, &seen, &reada, &nodes,
3448                                             &found_key);
3449                         free_extent_buffer(buf);
3450                 }
3451                 path.slots[0]++;
3452         }
3453         btrfs_release_path(root, &path);
3454         while(1) {
3455                 ret = run_next_block(root, bits, bits_nr, &last, &pending,
3456                                      &seen, &reada, &nodes, &extent_cache);
3457                 if (ret != 0)
3458                         break;
3459         }
3460         ret = check_extent_refs(trans, root, &extent_cache, repair);
3461
3462         if (repair) {
3463                 free_corrupt_blocks(root->fs_info);
3464                 root->fs_info->fsck_extent_cache = NULL;
3465                 root->fs_info->free_extent_hook = NULL;
3466                 root->fs_info->corrupt_blocks = NULL;
3467         }
3468
3469         return ret;
3470 }
3471
3472 static void print_usage(void)
3473 {
3474         fprintf(stderr, "usage: btrfsck dev\n");
3475         fprintf(stderr, "%s\n", BTRFS_BUILD_VERSION);
3476         exit(1);
3477 }
3478
3479 static struct option long_options[] = {
3480         { "super", 1, NULL, 's' },
3481         { "repair", 0, NULL, 0 },
3482         { "init-csum-tree", 0, NULL, 0 },
3483         { "init-extent-tree", 0, NULL, 0 },
3484         { 0, 0, 0, 0}
3485 };
3486
3487 int main(int ac, char **av)
3488 {
3489         struct cache_tree root_cache;
3490         struct btrfs_root *root;
3491         struct btrfs_fs_info *info;
3492         struct btrfs_trans_handle *trans = NULL;
3493         u64 bytenr = 0;
3494         int ret;
3495         int num;
3496         int repair = 0;
3497         int option_index = 0;
3498         int init_csum_tree = 0;
3499         int rw = 0;
3500
3501         while(1) {
3502                 int c;
3503                 c = getopt_long(ac, av, "s:", long_options,
3504                                 &option_index);
3505                 if (c < 0)
3506                         break;
3507                 switch(c) {
3508                         case 's':
3509                                 num = atol(optarg);
3510                                 bytenr = btrfs_sb_offset(num);
3511                                 printf("using SB copy %d, bytenr %llu\n", num,
3512                                        (unsigned long long)bytenr);
3513                                 break;
3514                         case '?':
3515                                 print_usage();
3516                 }
3517                 if (option_index == 1) {
3518                         printf("enabling repair mode\n");
3519                         repair = 1;
3520                         rw = 1;
3521                 } else if (option_index == 2) {
3522                         printf("Creating a new CRC tree\n");
3523                         init_csum_tree = 1;
3524                         rw = 1;
3525                 }
3526
3527         }
3528         ac = ac - optind;
3529
3530         if (ac != 1)
3531                 print_usage();
3532
3533         radix_tree_init();
3534         cache_tree_init(&root_cache);
3535
3536         if((ret = check_mounted(av[optind])) < 0) {
3537                 fprintf(stderr, "Could not check mount status: %s\n", strerror(-ret));
3538                 return ret;
3539         } else if(ret) {
3540                 fprintf(stderr, "%s is currently mounted. Aborting.\n", av[optind]);
3541                 return -EBUSY;
3542         }
3543
3544         info = open_ctree_fs_info(av[optind], bytenr, rw, 1);
3545
3546         if (info == NULL)
3547                 return 1;
3548
3549         if (!extent_buffer_uptodate(info->tree_root->node) ||
3550             !extent_buffer_uptodate(info->dev_root->node) ||
3551             !extent_buffer_uptodate(info->extent_root->node) ||
3552             !extent_buffer_uptodate(info->chunk_root->node)) {
3553                 fprintf(stderr, "Critical roots corrupted, unable to fsck the FS\n");
3554                 return -EIO;
3555         }
3556
3557         root = info->fs_root;
3558
3559         fprintf(stderr, "checking extents\n");
3560         if (rw)
3561                 trans = btrfs_start_transaction(root, 1);
3562
3563         if (init_csum_tree) {
3564                 fprintf(stderr, "Reinit crc root\n");
3565                 ret = btrfs_fsck_reinit_root(trans, info->csum_root);
3566                 if (ret) {
3567                         fprintf(stderr, "crc root initialization failed\n");
3568                         return -EIO;
3569                 }
3570                 goto out;
3571         }
3572         ret = check_extents(trans, root, repair);
3573         if (ret)
3574                 fprintf(stderr, "Errors found in extent allocation tree\n");
3575
3576         fprintf(stderr, "checking fs roots\n");
3577         ret = check_fs_roots(root, &root_cache);
3578         if (ret)
3579                 goto out;
3580
3581         fprintf(stderr, "checking root refs\n");
3582         ret = check_root_refs(root, &root_cache);
3583 out:
3584         free_root_recs(&root_cache);
3585         if (rw) {
3586                 ret = btrfs_commit_transaction(trans, root);
3587                 if (ret)
3588                         exit(1);
3589         }
3590         close_ctree(root);
3591
3592         if (found_old_backref) { /*
3593                  * there was a disk format change when mixed
3594                  * backref was in testing tree. The old format
3595                  * existed about one week.
3596                  */
3597                 printf("\n * Found old mixed backref format. "
3598                        "The old format is not supported! *"
3599                        "\n * Please mount the FS in readonly mode, "
3600                        "backup data and re-format the FS. *\n\n");
3601                 ret = 1;
3602         }
3603         printf("found %llu bytes used err is %d\n",
3604                (unsigned long long)bytes_used, ret);
3605         printf("total csum bytes: %llu\n",(unsigned long long)total_csum_bytes);
3606         printf("total tree bytes: %llu\n",
3607                (unsigned long long)total_btree_bytes);
3608         printf("total fs tree bytes: %llu\n",
3609                (unsigned long long)total_fs_tree_bytes);
3610         printf("btree space waste bytes: %llu\n",
3611                (unsigned long long)btree_space_waste);
3612         printf("file data blocks allocated: %llu\n referenced %llu\n",
3613                 (unsigned long long)data_bytes_allocated,
3614                 (unsigned long long)data_bytes_referenced);
3615         printf("%s\n", BTRFS_BUILD_VERSION);
3616         return ret;
3617 }