cmds-check.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/types.h>
  26 #include <sys/stat.h>
  27 #include <unistd.h>
  28 #include <getopt.h>
  29 #include <uuid/uuid.h>
  30 #include "kerncompat.h"
  31 #include "ctree.h"
  32 #include "volumes.h"
  33 #include "repair.h"
  34 #include "disk-io.h"
  35 #include "print-tree.h"
  36 #include "transaction.h"
  37 #include "list.h"
  38 #include "version.h"
  39 #include "utils.h"
  40 #include "commands.h"
  41 #include "free-space-cache.h"
  42
  43 static u64 bytes_used = 0;
  44 static u64 total_csum_bytes = 0;
  45 static u64 total_btree_bytes = 0;
  46 static u64 total_fs_tree_bytes = 0;
  47 static u64 total_extent_tree_bytes = 0;
  48 static u64 btree_space_waste = 0;
  49 static u64 data_bytes_allocated = 0;
  50 static u64 data_bytes_referenced = 0;
  51 static int found_old_backref = 0;
  52
  53 struct extent_backref {
  54         struct list_head list;
  55         unsigned int is_data:1;
  56         unsigned int found_extent_tree:1;
  57         unsigned int full_backref:1;
  58         unsigned int found_ref:1;
  59 };
  60
  61 struct data_backref {
  62         struct extent_backref node;
  63         union {
  64                 u64 parent;
  65                 u64 root;
  66         };
  67         u64 owner;
  68         u64 offset;
  69         u64 bytes;
  70         u32 num_refs;
  71         u32 found_ref;
  72 };
  73
  74 struct tree_backref {
  75         struct extent_backref node;
  76         union {
  77                 u64 parent;
  78                 u64 root;
  79         };
  80 };
  81
  82 struct extent_record {
  83         struct list_head backrefs;
  84         struct cache_extent cache;
  85         struct btrfs_disk_key parent_key;
  86         u64 start;
  87         u64 max_size;
  88         u64 nr;
  89         u64 refs;
  90         u64 extent_item_refs;
  91         u64 generation;
  92         u64 info_objectid;
  93         u8 info_level;
  94         unsigned int content_checked:1;
  95         unsigned int owner_ref_checked:1;
  96         unsigned int is_root:1;
  97         unsigned int metadata:1;
  98 };
  99
 100 struct inode_backref {
 101         struct list_head list;
 102         unsigned int found_dir_item:1;
 103         unsigned int found_dir_index:1;
 104         unsigned int found_inode_ref:1;
 105         unsigned int filetype:8;
 106         int errors;
 107         unsigned int ref_type;
 108         u64 dir;
 109         u64 index;
 110         u16 namelen;
 111         char name[0];
 112 };
 113
 114 #define REF_ERR_NO_DIR_ITEM             (1 << 0)
 115 #define REF_ERR_NO_DIR_INDEX            (1 << 1)
 116 #define REF_ERR_NO_INODE_REF            (1 << 2)
 117 #define REF_ERR_DUP_DIR_ITEM            (1 << 3)
 118 #define REF_ERR_DUP_DIR_INDEX           (1 << 4)
 119 #define REF_ERR_DUP_INODE_REF           (1 << 5)
 120 #define REF_ERR_INDEX_UNMATCH           (1 << 6)
 121 #define REF_ERR_FILETYPE_UNMATCH        (1 << 7)
 122 #define REF_ERR_NAME_TOO_LONG           (1 << 8) // 100
 123 #define REF_ERR_NO_ROOT_REF             (1 << 9)
 124 #define REF_ERR_NO_ROOT_BACKREF         (1 << 10)
 125 #define REF_ERR_DUP_ROOT_REF            (1 << 11)
 126 #define REF_ERR_DUP_ROOT_BACKREF        (1 << 12)
 127
 128 struct inode_record {
 129         struct list_head backrefs;
 130         unsigned int checked:1;
 131         unsigned int merging:1;
 132         unsigned int found_inode_item:1;
 133         unsigned int found_dir_item:1;
 134         unsigned int found_file_extent:1;
 135         unsigned int found_csum_item:1;
 136         unsigned int some_csum_missing:1;
 137         unsigned int nodatasum:1;
 138         int errors;
 139
 140         u64 ino;
 141         u32 nlink;
 142         u32 imode;
 143         u64 isize;
 144         u64 nbytes;
 145
 146         u32 found_link;
 147         u64 found_size;
 148         u64 extent_start;
 149         u64 extent_end;
 150         u64 first_extent_gap;
 151
 152         u32 refs;
 153 };
 154
 155 #define I_ERR_NO_INODE_ITEM             (1 << 0)
 156 #define I_ERR_NO_ORPHAN_ITEM            (1 << 1)
 157 #define I_ERR_DUP_INODE_ITEM            (1 << 2)
 158 #define I_ERR_DUP_DIR_INDEX             (1 << 3)
 159 #define I_ERR_ODD_DIR_ITEM              (1 << 4)
 160 #define I_ERR_ODD_FILE_EXTENT           (1 << 5)
 161 #define I_ERR_BAD_FILE_EXTENT           (1 << 6)
 162 #define I_ERR_FILE_EXTENT_OVERLAP       (1 << 7)
 163 #define I_ERR_FILE_EXTENT_DISCOUNT      (1 << 8) // 100
 164 #define I_ERR_DIR_ISIZE_WRONG           (1 << 9)
 165 #define I_ERR_FILE_NBYTES_WRONG         (1 << 10) // 400
 166 #define I_ERR_ODD_CSUM_ITEM             (1 << 11)
 167 #define I_ERR_SOME_CSUM_MISSING         (1 << 12)
 168 #define I_ERR_LINK_COUNT_WRONG          (1 << 13)
 169
 170 struct root_backref {
 171         struct list_head list;
 172         unsigned int found_dir_item:1;
 173         unsigned int found_dir_index:1;
 174         unsigned int found_back_ref:1;
 175         unsigned int found_forward_ref:1;
 176         unsigned int reachable:1;
 177         int errors;
 178         u64 ref_root;
 179         u64 dir;
 180         u64 index;
 181         u16 namelen;
 182         char name[0];
 183 };
 184
 185 struct root_record {
 186         struct list_head backrefs;
 187         struct cache_extent cache;
 188         unsigned int found_root_item:1;
 189         u64 objectid;
 190         u32 found_ref;
 191 };
 192
 193 struct ptr_node {
 194         struct cache_extent cache;
 195         void *data;
 196 };
 197
 198 struct shared_node {
 199         struct cache_extent cache;
 200         struct cache_tree root_cache;
 201         struct cache_tree inode_cache;
 202         struct inode_record *current;
 203         u32 refs;
 204 };
 205
 206 struct block_info {
 207         u64 start;
 208         u32 size;
 209 };
 210
 211 struct walk_control {
 212         struct cache_tree shared;
 213         struct shared_node *nodes[BTRFS_MAX_LEVEL];
 214         int active_node;
 215         int root_level;
 216 };
 217
 218 static u8 imode_to_type(u32 imode)
 219 {
 220 #define S_SHIFT 12
 221         static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
 222                 [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
 223                 [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
 224                 [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
 225                 [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
 226                 [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 227                 [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 228                 [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 229         };
 230
 231         return btrfs_type_by_mode[(imode & S_IFMT) >> S_SHIFT];
 232 #undef S_SHIFT
 233 }
 234
 235 static struct inode_record *clone_inode_rec(struct inode_record *orig_rec)
 236 {
 237         struct inode_record *rec;
 238         struct inode_backref *backref;
 239         struct inode_backref *orig;
 240         size_t size;
 241
 242         rec = malloc(sizeof(*rec));
 243         memcpy(rec, orig_rec, sizeof(*rec));
 244         rec->refs = 1;
 245         INIT_LIST_HEAD(&rec->backrefs);
 246
 247         list_for_each_entry(orig, &orig_rec->backrefs, list) {
 248                 size = sizeof(*orig) + orig->namelen + 1;
 249                 backref = malloc(size);
 250                 memcpy(backref, orig, size);
 251                 list_add_tail(&backref->list, &rec->backrefs);
 252         }
 253         return rec;
 254 }
 255
 256 static struct inode_record *get_inode_rec(struct cache_tree *inode_cache,
 257                                           u64 ino, int mod)
 258 {
 259         struct ptr_node *node;
 260         struct cache_extent *cache;
 261         struct inode_record *rec = NULL;
 262         int ret;
 263
 264         cache = find_cache_extent(inode_cache, ino, 1);
 265         if (cache) {
 266                 node = container_of(cache, struct ptr_node, cache);
 267                 rec = node->data;
 268                 if (mod && rec->refs > 1) {
 269                         node->data = clone_inode_rec(rec);
 270                         rec->refs--;
 271                         rec = node->data;
 272                 }
 273         } else if (mod) {
 274                 rec = calloc(1, sizeof(*rec));
 275                 rec->ino = ino;
 276                 rec->extent_start = (u64)-1;
 277                 rec->first_extent_gap = (u64)-1;
 278                 rec->refs = 1;
 279                 INIT_LIST_HEAD(&rec->backrefs);
 280
 281                 node = malloc(sizeof(*node));
 282                 node->cache.start = ino;
 283                 node->cache.size = 1;
 284                 node->data = rec;
 285
 286                 if (ino == BTRFS_FREE_INO_OBJECTID)
 287                         rec->found_link = 1;
 288
 289                 ret = insert_existing_cache_extent(inode_cache, &node->cache);
 290                 BUG_ON(ret);
 291         }
 292         return rec;
 293 }
 294
 295 static void free_inode_rec(struct inode_record *rec)
 296 {
 297         struct inode_backref *backref;
 298
 299         if (--rec->refs > 0)
 300                 return;
 301
 302         while (!list_empty(&rec->backrefs)) {
 303                 backref = list_entry(rec->backrefs.next,
 304                                      struct inode_backref, list);
 305                 list_del(&backref->list);
 306                 free(backref);
 307         }
 308         free(rec);
 309 }
 310
 311 static int can_free_inode_rec(struct inode_record *rec)
 312 {
 313         if (!rec->errors && rec->checked && rec->found_inode_item &&
 314             rec->nlink == rec->found_link && list_empty(&rec->backrefs))
 315                 return 1;
 316         return 0;
 317 }
 318
 319 static void maybe_free_inode_rec(struct cache_tree *inode_cache,
 320                                  struct inode_record *rec)
 321 {
 322         struct cache_extent *cache;
 323         struct inode_backref *tmp, *backref;
 324         struct ptr_node *node;
 325         unsigned char filetype;
 326
 327         if (!rec->found_inode_item)
 328                 return;
 329
 330         filetype = imode_to_type(rec->imode);
 331         list_for_each_entry_safe(backref, tmp, &rec->backrefs, list) {
 332                 if (backref->found_dir_item && backref->found_dir_index) {
 333                         if (backref->filetype != filetype)
 334                                 backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 335                         if (!backref->errors && backref->found_inode_ref) {
 336                                 list_del(&backref->list);
 337                                 free(backref);
 338                         }
 339                 }
 340         }
 341
 342         if (!rec->checked || rec->merging)
 343                 return;
 344
 345         if (S_ISDIR(rec->imode)) {
 346                 if (rec->found_size != rec->isize)
 347                         rec->errors |= I_ERR_DIR_ISIZE_WRONG;
 348                 if (rec->found_file_extent)
 349                         rec->errors |= I_ERR_ODD_FILE_EXTENT;
 350         } else if (S_ISREG(rec->imode) || S_ISLNK(rec->imode)) {
 351                 if (rec->found_dir_item)
 352                         rec->errors |= I_ERR_ODD_DIR_ITEM;
 353                 if (rec->found_size != rec->nbytes)
 354                         rec->errors |= I_ERR_FILE_NBYTES_WRONG;
 355                 if (rec->extent_start == (u64)-1 || rec->extent_start > 0)
 356                         rec->first_extent_gap = 0;
 357                 if (rec->nlink > 0 && (rec->extent_end < rec->isize ||
 358                     rec->first_extent_gap < rec->isize))
 359                         rec->errors |= I_ERR_FILE_EXTENT_DISCOUNT;
 360         }
 361
 362         if (S_ISREG(rec->imode) || S_ISLNK(rec->imode)) {
 363                 if (rec->found_csum_item && rec->nodatasum)
 364                         rec->errors |= I_ERR_ODD_CSUM_ITEM;
 365                 if (rec->some_csum_missing && !rec->nodatasum)
 366                         rec->errors |= I_ERR_SOME_CSUM_MISSING;
 367         }
 368
 369         BUG_ON(rec->refs != 1);
 370         if (can_free_inode_rec(rec)) {
 371                 cache = find_cache_extent(inode_cache, rec->ino, 1);
 372                 node = container_of(cache, struct ptr_node, cache);
 373                 BUG_ON(node->data != rec);
 374                 remove_cache_extent(inode_cache, &node->cache);
 375                 free(node);
 376                 free_inode_rec(rec);
 377         }
 378 }
 379
 380 static int check_orphan_item(struct btrfs_root *root, u64 ino)
 381 {
 382         struct btrfs_path path;
 383         struct btrfs_key key;
 384         int ret;
 385
 386         key.objectid = BTRFS_ORPHAN_OBJECTID;
 387         key.type = BTRFS_ORPHAN_ITEM_KEY;
 388         key.offset = ino;
 389
 390         btrfs_init_path(&path);
 391         ret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
 392         btrfs_release_path(root, &path);
 393         if (ret > 0)
 394                 ret = -ENOENT;
 395         return ret;
 396 }
 397
 398 static int process_inode_item(struct extent_buffer *eb,
 399                               int slot, struct btrfs_key *key,
 400                               struct shared_node *active_node)
 401 {
 402         struct inode_record *rec;
 403         struct btrfs_inode_item *item;
 404
 405         rec = active_node->current;
 406         BUG_ON(rec->ino != key->objectid || rec->refs > 1);
 407         if (rec->found_inode_item) {
 408                 rec->errors |= I_ERR_DUP_INODE_ITEM;
 409                 return 1;
 410         }
 411         item = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
 412         rec->nlink = btrfs_inode_nlink(eb, item);
 413         rec->isize = btrfs_inode_size(eb, item);
 414         rec->nbytes = btrfs_inode_nbytes(eb, item);
 415         rec->imode = btrfs_inode_mode(eb, item);
 416         if (btrfs_inode_flags(eb, item) & BTRFS_INODE_NODATASUM)
 417                 rec->nodatasum = 1;
 418         rec->found_inode_item = 1;
 419         if (rec->nlink == 0)
 420                 rec->errors |= I_ERR_NO_ORPHAN_ITEM;
 421         maybe_free_inode_rec(&active_node->inode_cache, rec);
 422         return 0;
 423 }
 424
 425 static struct inode_backref *get_inode_backref(struct inode_record *rec,
 426                                                 const char *name,
 427                                                 int namelen, u64 dir)
 428 {
 429         struct inode_backref *backref;
 430
 431         list_for_each_entry(backref, &rec->backrefs, list) {
 432                 if (backref->dir != dir || backref->namelen != namelen)
 433                         continue;
 434                 if (memcmp(name, backref->name, namelen))
 435                         continue;
 436                 return backref;
 437         }
 438
 439         backref = malloc(sizeof(*backref) + namelen + 1);
 440         memset(backref, 0, sizeof(*backref));
 441         backref->dir = dir;
 442         backref->namelen = namelen;
 443         memcpy(backref->name, name, namelen);
 444         backref->name[namelen] = '\0';
 445         list_add_tail(&backref->list, &rec->backrefs);
 446         return backref;
 447 }
 448
 449 static int add_inode_backref(struct cache_tree *inode_cache,
 450                              u64 ino, u64 dir, u64 index,
 451                              const char *name, int namelen,
 452                              int filetype, int itemtype, int errors)
 453 {
 454         struct inode_record *rec;
 455         struct inode_backref *backref;
 456
 457         rec = get_inode_rec(inode_cache, ino, 1);
 458         backref = get_inode_backref(rec, name, namelen, dir);
 459         if (errors)
 460                 backref->errors |= errors;
 461         if (itemtype == BTRFS_DIR_INDEX_KEY) {
 462                 if (backref->found_dir_index)
 463                         backref->errors |= REF_ERR_DUP_DIR_INDEX;
 464                 if (backref->found_inode_ref && backref->index != index)
 465                         backref->errors |= REF_ERR_INDEX_UNMATCH;
 466                 if (backref->found_dir_item && backref->filetype != filetype)
 467                         backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 468
 469                 backref->index = index;
 470                 backref->filetype = filetype;
 471                 backref->found_dir_index = 1;
 472         } else if (itemtype == BTRFS_DIR_ITEM_KEY) {
 473                 rec->found_link++;
 474                 if (backref->found_dir_item)
 475                         backref->errors |= REF_ERR_DUP_DIR_ITEM;
 476                 if (backref->found_dir_index && backref->filetype != filetype)
 477                         backref->errors |= REF_ERR_FILETYPE_UNMATCH;
 478
 479                 backref->filetype = filetype;
 480                 backref->found_dir_item = 1;
 481         } else if ((itemtype == BTRFS_INODE_REF_KEY) ||
 482                    (itemtype == BTRFS_INODE_EXTREF_KEY)) {
 483                 if (backref->found_inode_ref)
 484                         backref->errors |= REF_ERR_DUP_INODE_REF;
 485                 if (backref->found_dir_index && backref->index != index)
 486                         backref->errors |= REF_ERR_INDEX_UNMATCH;
 487
 488                 backref->ref_type = itemtype;
 489                 backref->index = index;
 490                 backref->found_inode_ref = 1;
 491         } else {
 492                 BUG_ON(1);
 493         }
 494
 495         maybe_free_inode_rec(inode_cache, rec);
 496         return 0;
 497 }
 498
 499 static int merge_inode_recs(struct inode_record *src, struct inode_record *dst,
 500                             struct cache_tree *dst_cache)
 501 {
 502         struct inode_backref *backref;
 503         u32 dir_count = 0;
 504
 505         dst->merging = 1;
 506         list_for_each_entry(backref, &src->backrefs, list) {
 507                 if (backref->found_dir_index) {
 508                         add_inode_backref(dst_cache, dst->ino, backref->dir,
 509                                         backref->index, backref->name,
 510                                         backref->namelen, backref->filetype,
 511                                         BTRFS_DIR_INDEX_KEY, backref->errors);
 512                 }
 513                 if (backref->found_dir_item) {
 514                         dir_count++;
 515                         add_inode_backref(dst_cache, dst->ino,
 516                                         backref->dir, 0, backref->name,
 517                                         backref->namelen, backref->filetype,
 518                                         BTRFS_DIR_ITEM_KEY, backref->errors);
 519                 }
 520                 if (backref->found_inode_ref) {
 521                         add_inode_backref(dst_cache, dst->ino,
 522                                         backref->dir, backref->index,
 523                                         backref->name, backref->namelen, 0,
 524                                         backref->ref_type, backref->errors);
 525                 }
 526         }
 527
 528         if (src->found_dir_item)
 529                 dst->found_dir_item = 1;
 530         if (src->found_file_extent)
 531                 dst->found_file_extent = 1;
 532         if (src->found_csum_item)
 533                 dst->found_csum_item = 1;
 534         if (src->some_csum_missing)
 535                 dst->some_csum_missing = 1;
 536         if (dst->first_extent_gap > src->first_extent_gap)
 537                 dst->first_extent_gap = src->first_extent_gap;
 538
 539         BUG_ON(src->found_link < dir_count);
 540         dst->found_link += src->found_link - dir_count;
 541         dst->found_size += src->found_size;
 542         if (src->extent_start != (u64)-1) {
 543                 if (dst->extent_start == (u64)-1) {
 544                         dst->extent_start = src->extent_start;
 545                         dst->extent_end = src->extent_end;
 546                 } else {
 547                         if (dst->extent_end > src->extent_start)
 548                                 dst->errors |= I_ERR_FILE_EXTENT_OVERLAP;
 549                         else if (dst->extent_end < src->extent_start &&
 550                                  dst->extent_end < dst->first_extent_gap)
 551                                 dst->first_extent_gap = dst->extent_end;
 552                         if (dst->extent_end < src->extent_end)
 553                                 dst->extent_end = src->extent_end;
 554                 }
 555         }
 556
 557         dst->errors |= src->errors;
 558         if (src->found_inode_item) {
 559                 if (!dst->found_inode_item) {
 560                         dst->nlink = src->nlink;
 561                         dst->isize = src->isize;
 562                         dst->nbytes = src->nbytes;
 563                         dst->imode = src->imode;
 564                         dst->nodatasum = src->nodatasum;
 565                         dst->found_inode_item = 1;
 566                 } else {
 567                         dst->errors |= I_ERR_DUP_INODE_ITEM;
 568                 }
 569         }
 570         dst->merging = 0;
 571
 572         return 0;
 573 }
 574
 575 static int splice_shared_node(struct shared_node *src_node,
 576                               struct shared_node *dst_node)
 577 {
 578         struct cache_extent *cache;
 579         struct ptr_node *node, *ins;
 580         struct cache_tree *src, *dst;
 581         struct inode_record *rec, *conflict;
 582         u64 current_ino = 0;
 583         int splice = 0;
 584         int ret;
 585
 586         if (--src_node->refs == 0)
 587                 splice = 1;
 588         if (src_node->current)
 589                 current_ino = src_node->current->ino;
 590
 591         src = &src_node->root_cache;
 592         dst = &dst_node->root_cache;
 593 again:
 594         cache = find_first_cache_extent(src, 0);
 595         while (cache) {
 596                 node = container_of(cache, struct ptr_node, cache);
 597                 rec = node->data;
 598                 cache = next_cache_extent(cache);
 599
 600                 if (splice) {
 601                         remove_cache_extent(src, &node->cache);
 602                         ins = node;
 603                 } else {
 604                         ins = malloc(sizeof(*ins));
 605                         ins->cache.start = node->cache.start;
 606                         ins->cache.size = node->cache.size;
 607                         ins->data = rec;
 608                         rec->refs++;
 609                 }
 610                 ret = insert_existing_cache_extent(dst, &ins->cache);
 611                 if (ret == -EEXIST) {
 612                         conflict = get_inode_rec(dst, rec->ino, 1);
 613                         merge_inode_recs(rec, conflict, dst);
 614                         if (rec->checked) {
 615                                 conflict->checked = 1;
 616                                 if (dst_node->current == conflict)
 617                                         dst_node->current = NULL;
 618                         }
 619                         maybe_free_inode_rec(dst, conflict);
 620                         free_inode_rec(rec);
 621                         free(ins);
 622                 } else {
 623                         BUG_ON(ret);
 624                 }
 625         }
 626
 627         if (src == &src_node->root_cache) {
 628                 src = &src_node->inode_cache;
 629                 dst = &dst_node->inode_cache;
 630                 goto again;
 631         }
 632
 633         if (current_ino > 0 && (!dst_node->current ||
 634             current_ino > dst_node->current->ino)) {
 635                 if (dst_node->current) {
 636                         dst_node->current->checked = 1;
 637                         maybe_free_inode_rec(dst, dst_node->current);
 638                 }
 639                 dst_node->current = get_inode_rec(dst, current_ino, 1);
 640         }
 641         return 0;
 642 }
 643
 644 static void free_inode_recs(struct cache_tree *inode_cache)
 645 {
 646         struct cache_extent *cache;
 647         struct ptr_node *node;
 648         struct inode_record *rec;
 649
 650         while (1) {
 651                 cache = find_first_cache_extent(inode_cache, 0);
 652                 if (!cache)
 653                         break;
 654                 node = container_of(cache, struct ptr_node, cache);
 655                 rec = node->data;
 656                 remove_cache_extent(inode_cache, &node->cache);
 657                 free(node);
 658                 free_inode_rec(rec);
 659         }
 660 }
 661
 662 static struct shared_node *find_shared_node(struct cache_tree *shared,
 663                                             u64 bytenr)
 664 {
 665         struct cache_extent *cache;
 666         struct shared_node *node;
 667
 668         cache = find_cache_extent(shared, bytenr, 1);
 669         if (cache) {
 670                 node = container_of(cache, struct shared_node, cache);
 671                 return node;
 672         }
 673         return NULL;
 674 }
 675
 676 static int add_shared_node(struct cache_tree *shared, u64 bytenr, u32 refs)
 677 {
 678         int ret;
 679         struct shared_node *node;
 680
 681         node = calloc(1, sizeof(*node));
 682         node->cache.start = bytenr;
 683         node->cache.size = 1;
 684         cache_tree_init(&node->root_cache);
 685         cache_tree_init(&node->inode_cache);
 686         node->refs = refs;
 687
 688         ret = insert_existing_cache_extent(shared, &node->cache);
 689         BUG_ON(ret);
 690         return 0;
 691 }
 692
 693 static int enter_shared_node(struct btrfs_root *root, u64 bytenr, u32 refs,
 694                              struct walk_control *wc, int level)
 695 {
 696         struct shared_node *node;
 697         struct shared_node *dest;
 698
 699         if (level == wc->active_node)
 700                 return 0;
 701
 702         BUG_ON(wc->active_node <= level);
 703         node = find_shared_node(&wc->shared, bytenr);
 704         if (!node) {
 705                 add_shared_node(&wc->shared, bytenr, refs);
 706                 node = find_shared_node(&wc->shared, bytenr);
 707                 wc->nodes[level] = node;
 708                 wc->active_node = level;
 709                 return 0;
 710         }
 711
 712         if (wc->root_level == wc->active_node &&
 713             btrfs_root_refs(&root->root_item) == 0) {
 714                 if (--node->refs == 0) {
 715                         free_inode_recs(&node->root_cache);
 716                         free_inode_recs(&node->inode_cache);
 717                         remove_cache_extent(&wc->shared, &node->cache);
 718                         free(node);
 719                 }
 720                 return 1;
 721         }
 722
 723         dest = wc->nodes[wc->active_node];
 724         splice_shared_node(node, dest);
 725         if (node->refs == 0) {
 726                 remove_cache_extent(&wc->shared, &node->cache);
 727                 free(node);
 728         }
 729         return 1;
 730 }
 731
 732 static int leave_shared_node(struct btrfs_root *root,
 733                              struct walk_control *wc, int level)
 734 {
 735         struct shared_node *node;
 736         struct shared_node *dest;
 737         int i;
 738
 739         if (level == wc->root_level)
 740                 return 0;
 741
 742         for (i = level + 1; i < BTRFS_MAX_LEVEL; i++) {
 743                 if (wc->nodes[i])
 744                         break;
 745         }
 746         BUG_ON(i >= BTRFS_MAX_LEVEL);
 747
 748         node = wc->nodes[wc->active_node];
 749         wc->nodes[wc->active_node] = NULL;
 750         wc->active_node = i;
 751
 752         dest = wc->nodes[wc->active_node];
 753         if (wc->active_node < wc->root_level ||
 754             btrfs_root_refs(&root->root_item) > 0) {
 755                 BUG_ON(node->refs <= 1);
 756                 splice_shared_node(node, dest);
 757         } else {
 758                 BUG_ON(node->refs < 2);
 759                 node->refs--;
 760         }
 761         return 0;
 762 }
 763
 764 static int is_child_root(struct btrfs_root *root, u64 parent_root_id,
 765                          u64 child_root_id)
 766 {
 767         struct btrfs_path path;
 768         struct btrfs_key key;
 769         struct extent_buffer *leaf;
 770         int has_parent = 0;
 771         int ret;
 772
 773         btrfs_init_path(&path);
 774
 775         key.objectid = parent_root_id;
 776         key.type = BTRFS_ROOT_REF_KEY;
 777         key.offset = child_root_id;
 778         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, &path,
 779                                 0, 0);
 780         BUG_ON(ret < 0);
 781         btrfs_release_path(root, &path);
 782         if (!ret)
 783                 return 1;
 784
 785         key.objectid = child_root_id;
 786         key.type = BTRFS_ROOT_BACKREF_KEY;
 787         key.offset = 0;
 788         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, &path,
 789                                 0, 0);
 790         BUG_ON(ret <= 0);
 791
 792         while (1) {
 793                 leaf = path.nodes[0];
 794                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
 795                         ret = btrfs_next_leaf(root->fs_info->tree_root, &path);
 796                         BUG_ON(ret < 0);
 797
 798                         if (ret > 0)
 799                                 break;
 800                         leaf = path.nodes[0];
 801                 }
 802
 803                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
 804                 if (key.objectid != child_root_id ||
 805                     key.type != BTRFS_ROOT_BACKREF_KEY)
 806                         break;
 807
 808                 has_parent = 1;
 809
 810                 if (key.offset == parent_root_id) {
 811                         btrfs_release_path(root, &path);
 812                         return 1;
 813                 }
 814
 815                 path.slots[0]++;
 816         }
 817
 818         btrfs_release_path(root, &path);
 819         return has_parent? 0 : -1;
 820 }
 821
 822 static int process_dir_item(struct btrfs_root *root,
 823                             struct extent_buffer *eb,
 824                             int slot, struct btrfs_key *key,
 825                             struct shared_node *active_node)
 826 {
 827         u32 total;
 828         u32 cur = 0;
 829         u32 len;
 830         u32 name_len;
 831         u32 data_len;
 832         int error;
 833         int nritems = 0;
 834         int filetype;
 835         struct btrfs_dir_item *di;
 836         struct inode_record *rec;
 837         struct cache_tree *root_cache;
 838         struct cache_tree *inode_cache;
 839         struct btrfs_key location;
 840         char namebuf[BTRFS_NAME_LEN];
 841
 842         root_cache = &active_node->root_cache;
 843         inode_cache = &active_node->inode_cache;
 844         rec = active_node->current;
 845         rec->found_dir_item = 1;
 846
 847         di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
 848         total = btrfs_item_size_nr(eb, slot);
 849         while (cur < total) {
 850                 nritems++;
 851                 btrfs_dir_item_key_to_cpu(eb, di, &location);
 852                 name_len = btrfs_dir_name_len(eb, di);
 853                 data_len = btrfs_dir_data_len(eb, di);
 854                 filetype = btrfs_dir_type(eb, di);
 855
 856                 rec->found_size += name_len;
 857                 if (name_len <= BTRFS_NAME_LEN) {
 858                         len = name_len;
 859                         error = 0;
 860                 } else {
 861                         len = BTRFS_NAME_LEN;
 862                         error = REF_ERR_NAME_TOO_LONG;
 863                 }
 864                 read_extent_buffer(eb, namebuf, (unsigned long)(di + 1), len);
 865
 866                 if (location.type == BTRFS_INODE_ITEM_KEY) {
 867                         add_inode_backref(inode_cache, location.objectid,
 868                                           key->objectid, key->offset, namebuf,
 869                                           len, filetype, key->type, error);
 870                 } else if (location.type == BTRFS_ROOT_ITEM_KEY) {
 871                         add_inode_backref(root_cache, location.objectid,
 872                                           key->objectid, key->offset,
 873                                           namebuf, len, filetype,
 874                                           key->type, error);
 875                 } else {
 876                         fprintf(stderr, "warning line %d\n", __LINE__);
 877                 }
 878
 879                 len = sizeof(*di) + name_len + data_len;
 880                 di = (struct btrfs_dir_item *)((char *)di + len);
 881                 cur += len;
 882         }
 883         if (key->type == BTRFS_DIR_INDEX_KEY && nritems > 1)
 884                 rec->errors |= I_ERR_DUP_DIR_INDEX;
 885
 886         return 0;
 887 }
 888
 889 static int process_inode_ref(struct extent_buffer *eb,
 890                              int slot, struct btrfs_key *key,
 891                              struct shared_node *active_node)
 892 {
 893         u32 total;
 894         u32 cur = 0;
 895         u32 len;
 896         u32 name_len;
 897         u64 index;
 898         int error;
 899         struct cache_tree *inode_cache;
 900         struct btrfs_inode_ref *ref;
 901         char namebuf[BTRFS_NAME_LEN];
 902
 903         inode_cache = &active_node->inode_cache;
 904
 905         ref = btrfs_item_ptr(eb, slot, struct btrfs_inode_ref);
 906         total = btrfs_item_size_nr(eb, slot);
 907         while (cur < total) {
 908                 name_len = btrfs_inode_ref_name_len(eb, ref);
 909                 index = btrfs_inode_ref_index(eb, ref);
 910                 if (name_len <= BTRFS_NAME_LEN) {
 911                         len = name_len;
 912                         error = 0;
 913                 } else {
 914                         len = BTRFS_NAME_LEN;
 915                         error = REF_ERR_NAME_TOO_LONG;
 916                 }
 917                 read_extent_buffer(eb, namebuf, (unsigned long)(ref + 1), len);
 918                 add_inode_backref(inode_cache, key->objectid, key->offset,
 919                                   index, namebuf, len, 0, key->type, error);
 920
 921                 len = sizeof(*ref) + name_len;
 922                 ref = (struct btrfs_inode_ref *)((char *)ref + len);
 923                 cur += len;
 924         }
 925         return 0;
 926 }
 927
 928 static int process_inode_extref(struct extent_buffer *eb,
 929                                 int slot, struct btrfs_key *key,
 930                                 struct shared_node *active_node)
 931 {
 932         u32 total;
 933         u32 cur = 0;
 934         u32 len;
 935         u32 name_len;
 936         u64 index;
 937         u64 parent;
 938         int error;
 939         struct cache_tree *inode_cache;
 940         struct btrfs_inode_extref *extref;
 941         char namebuf[BTRFS_NAME_LEN];
 942
 943         inode_cache = &active_node->inode_cache;
 944
 945         extref = btrfs_item_ptr(eb, slot, struct btrfs_inode_extref);
 946         total = btrfs_item_size_nr(eb, slot);
 947         while (cur < total) {
 948                 name_len = btrfs_inode_extref_name_len(eb, extref);
 949                 index = btrfs_inode_extref_index(eb, extref);
 950                 parent = btrfs_inode_extref_parent(eb, extref);
 951                 if (name_len <= BTRFS_NAME_LEN) {
 952                         len = name_len;
 953                         error = 0;
 954                 } else {
 955                         len = BTRFS_NAME_LEN;
 956                         error = REF_ERR_NAME_TOO_LONG;
 957                 }
 958                 read_extent_buffer(eb, namebuf,
 959                                    (unsigned long)(extref + 1), len);
 960                 add_inode_backref(inode_cache, key->objectid, parent,
 961                                   index, namebuf, len, 0, key->type, error);
 962
 963                 len = sizeof(*extref) + name_len;
 964                 extref = (struct btrfs_inode_extref *)((char *)extref + len);
 965                 cur += len;
 966         }
 967         return 0;
 968
 969 }
 970
 971 static u64 count_csum_range(struct btrfs_root *root, u64 start, u64 len)
 972 {
 973         struct btrfs_key key;
 974         struct btrfs_path path;
 975         struct extent_buffer *leaf;
 976         int ret ;
 977         size_t size;
 978         u64 found = 0;
 979         u64 csum_end;
 980         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
 981
 982         btrfs_init_path(&path);
 983
 984         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
 985         key.offset = start;
 986         key.type = BTRFS_EXTENT_CSUM_KEY;
 987
 988         ret = btrfs_search_slot(NULL, root->fs_info->csum_root,
 989                                 &key, &path, 0, 0);
 990         BUG_ON(ret < 0);
 991         if (ret > 0 && path.slots[0] > 0) {
 992                 leaf = path.nodes[0];
 993                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0] - 1);
 994                 if (key.objectid == BTRFS_EXTENT_CSUM_OBJECTID &&
 995                     key.type == BTRFS_EXTENT_CSUM_KEY)
 996                         path.slots[0]--;
 997         }
 998
 999         while (len > 0) {
1000                 leaf = path.nodes[0];
1001                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1002                         ret = btrfs_next_leaf(root->fs_info->csum_root, &path);
1003                         BUG_ON(ret < 0);
1004                         if (ret > 0)
1005                                 break;
1006                         leaf = path.nodes[0];
1007                 }
1008
1009                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1010                 if (key.objectid != BTRFS_EXTENT_CSUM_OBJECTID ||
1011                     key.type != BTRFS_EXTENT_CSUM_KEY)
1012                         break;
1013
1014                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1015                 if (key.offset >= start + len)
1016                         break;
1017
1018                 if (key.offset > start)
1019                         start = key.offset;
1020
1021                 size = btrfs_item_size_nr(leaf, path.slots[0]);
1022                 csum_end = key.offset + (size / csum_size) * root->sectorsize;
1023                 if (csum_end > start) {
1024                         size = min(csum_end - start, len);
1025                         len -= size;
1026                         start += size;
1027                         found += size;
1028                 }
1029
1030                 path.slots[0]++;
1031         }
1032         btrfs_release_path(root->fs_info->csum_root, &path);
1033         return found;
1034 }
1035
1036 static int process_file_extent(struct btrfs_root *root,
1037                                 struct extent_buffer *eb,
1038                                 int slot, struct btrfs_key *key,
1039                                 struct shared_node *active_node)
1040 {
1041         struct inode_record *rec;
1042         struct btrfs_file_extent_item *fi;
1043         u64 num_bytes = 0;
1044         u64 disk_bytenr = 0;
1045         u64 extent_offset = 0;
1046         u64 mask = root->sectorsize - 1;
1047         int extent_type;
1048
1049         rec = active_node->current;
1050         BUG_ON(rec->ino != key->objectid || rec->refs > 1);
1051         rec->found_file_extent = 1;
1052
1053         if (rec->extent_start == (u64)-1) {
1054                 rec->extent_start = key->offset;
1055                 rec->extent_end = key->offset;
1056         }
1057
1058         if (rec->extent_end > key->offset)
1059                 rec->errors |= I_ERR_FILE_EXTENT_OVERLAP;
1060         else if (rec->extent_end < key->offset &&
1061                  rec->extent_end < rec->first_extent_gap)
1062                 rec->first_extent_gap = rec->extent_end;
1063
1064         fi = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
1065         extent_type = btrfs_file_extent_type(eb, fi);
1066
1067         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1068                 num_bytes = btrfs_file_extent_inline_len(eb, fi);
1069                 if (num_bytes == 0)
1070                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1071                 rec->found_size += num_bytes;
1072                 num_bytes = (num_bytes + mask) & ~mask;
1073         } else if (extent_type == BTRFS_FILE_EXTENT_REG ||
1074                    extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1075                 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1076                 disk_bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1077                 extent_offset = btrfs_file_extent_offset(eb, fi);
1078                 if (num_bytes == 0 || (num_bytes & mask))
1079                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1080                 if (num_bytes + extent_offset >
1081                     btrfs_file_extent_ram_bytes(eb, fi))
1082                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1083                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC &&
1084                     (btrfs_file_extent_compression(eb, fi) ||
1085                      btrfs_file_extent_encryption(eb, fi) ||
1086                      btrfs_file_extent_other_encoding(eb, fi)))
1087                         rec->errors |= I_ERR_BAD_FILE_EXTENT;
1088                 if (disk_bytenr > 0)
1089                         rec->found_size += num_bytes;
1090         } else {
1091                 rec->errors |= I_ERR_BAD_FILE_EXTENT;
1092         }
1093         rec->extent_end = key->offset + num_bytes;
1094
1095         if (disk_bytenr > 0) {
1096                 u64 found;
1097                 if (btrfs_file_extent_compression(eb, fi))
1098                         num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1099                 else
1100                         disk_bytenr += extent_offset;
1101
1102                 found = count_csum_range(root, disk_bytenr, num_bytes);
1103                 if (extent_type == BTRFS_FILE_EXTENT_REG) {
1104                         if (found > 0)
1105                                 rec->found_csum_item = 1;
1106                         if (found < num_bytes)
1107                                 rec->some_csum_missing = 1;
1108                 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1109                         if (found > 0)
1110                                 rec->errors |= I_ERR_ODD_CSUM_ITEM;
1111                 }
1112         }
1113         return 0;
1114 }
1115
1116 static int process_one_leaf(struct btrfs_root *root, struct extent_buffer *eb,
1117                             struct walk_control *wc)
1118 {
1119         struct btrfs_key key;
1120         u32 nritems;
1121         int i;
1122         int ret = 0;
1123         struct cache_tree *inode_cache;
1124         struct shared_node *active_node;
1125
1126         if (wc->root_level == wc->active_node &&
1127             btrfs_root_refs(&root->root_item) == 0)
1128                 return 0;
1129
1130         active_node = wc->nodes[wc->active_node];
1131         inode_cache = &active_node->inode_cache;
1132         nritems = btrfs_header_nritems(eb);
1133         for (i = 0; i < nritems; i++) {
1134                 btrfs_item_key_to_cpu(eb, &key, i);
1135
1136                 if (key.objectid == BTRFS_FREE_SPACE_OBJECTID)
1137                         continue;
1138
1139                 if (active_node->current == NULL ||
1140                     active_node->current->ino < key.objectid) {
1141                         if (active_node->current) {
1142                                 active_node->current->checked = 1;
1143                                 maybe_free_inode_rec(inode_cache,
1144                                                      active_node->current);
1145                         }
1146                         active_node->current = get_inode_rec(inode_cache,
1147                                                              key.objectid, 1);
1148                 }
1149                 switch (key.type) {
1150                 case BTRFS_DIR_ITEM_KEY:
1151                 case BTRFS_DIR_INDEX_KEY:
1152                         ret = process_dir_item(root, eb, i, &key, active_node);
1153                         break;
1154                 case BTRFS_INODE_REF_KEY:
1155                         ret = process_inode_ref(eb, i, &key, active_node);
1156                         break;
1157                 case BTRFS_INODE_EXTREF_KEY:
1158                         ret = process_inode_extref(eb, i, &key, active_node);
1159                         break;
1160                 case BTRFS_INODE_ITEM_KEY:
1161                         ret = process_inode_item(eb, i, &key, active_node);
1162                         break;
1163                 case BTRFS_EXTENT_DATA_KEY:
1164                         ret = process_file_extent(root, eb, i, &key,
1165                                                   active_node);
1166                         break;
1167                 default:
1168                         break;
1169                 };
1170         }
1171         return ret;
1172 }
1173
1174 static void reada_walk_down(struct btrfs_root *root,
1175                             struct extent_buffer *node, int slot)
1176 {
1177         u64 bytenr;
1178         u64 ptr_gen;
1179         u32 nritems;
1180         u32 blocksize;
1181         int i;
1182         int ret;
1183         int level;
1184
1185         level = btrfs_header_level(node);
1186         if (level != 1)
1187                 return;
1188
1189         nritems = btrfs_header_nritems(node);
1190         blocksize = btrfs_level_size(root, level - 1);
1191         for (i = slot; i < nritems; i++) {
1192                 bytenr = btrfs_node_blockptr(node, i);
1193                 ptr_gen = btrfs_node_ptr_generation(node, i);
1194                 ret = readahead_tree_block(root, bytenr, blocksize, ptr_gen);
1195                 if (ret)
1196                         break;
1197         }
1198 }
1199
1200 static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
1201                           struct walk_control *wc, int *level)
1202 {
1203         u64 bytenr;
1204         u64 ptr_gen;
1205         struct extent_buffer *next;
1206         struct extent_buffer *cur;
1207         u32 blocksize;
1208         int ret;
1209         u64 refs;
1210
1211         WARN_ON(*level < 0);
1212         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1213         ret = btrfs_lookup_extent_info(NULL, root,
1214                                        path->nodes[*level]->start,
1215                                        *level, 1, &refs, NULL);
1216         if (ret < 0)
1217                 goto out;
1218
1219         if (refs > 1) {
1220                 ret = enter_shared_node(root, path->nodes[*level]->start,
1221                                         refs, wc, *level);
1222                 if (ret > 0)
1223                         goto out;
1224         }
1225
1226         while (*level >= 0) {
1227                 WARN_ON(*level < 0);
1228                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1229                 cur = path->nodes[*level];
1230
1231                 if (btrfs_header_level(cur) != *level)
1232                         WARN_ON(1);
1233
1234                 if (path->slots[*level] >= btrfs_header_nritems(cur))
1235                         break;
1236                 if (*level == 0) {
1237                         ret = process_one_leaf(root, cur, wc);
1238                         break;
1239                 }
1240                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1241                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1242                 blocksize = btrfs_level_size(root, *level - 1);
1243                 ret = btrfs_lookup_extent_info(NULL, root, bytenr, *level - 1,
1244                                                1, &refs, NULL);
1245                 if (ret < 0)
1246                         refs = 0;
1247
1248                 if (refs > 1) {
1249                         ret = enter_shared_node(root, bytenr, refs,
1250                                                 wc, *level - 1);
1251                         if (ret > 0) {
1252                                 path->slots[*level]++;
1253                                 continue;
1254                         }
1255                 }
1256
1257                 next = btrfs_find_tree_block(root, bytenr, blocksize);
1258                 if (!next || !btrfs_buffer_uptodate(next, ptr_gen)) {
1259                         free_extent_buffer(next);
1260                         reada_walk_down(root, cur, path->slots[*level]);
1261                         next = read_tree_block(root, bytenr, blocksize,
1262                                                ptr_gen);
1263                 }
1264
1265                 *level = *level - 1;
1266                 free_extent_buffer(path->nodes[*level]);
1267                 path->nodes[*level] = next;
1268                 path->slots[*level] = 0;
1269         }
1270 out:
1271         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
1272         return 0;
1273 }
1274
1275 static int walk_up_tree(struct btrfs_root *root, struct btrfs_path *path,
1276                         struct walk_control *wc, int *level)
1277 {
1278         int i;
1279         struct extent_buffer *leaf;
1280
1281         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
1282                 leaf = path->nodes[i];
1283                 if (path->slots[i] + 1 < btrfs_header_nritems(leaf)) {
1284                         path->slots[i]++;
1285                         *level = i;
1286                         return 0;
1287                 } else {
1288                         free_extent_buffer(path->nodes[*level]);
1289                         path->nodes[*level] = NULL;
1290                         BUG_ON(*level > wc->active_node);
1291                         if (*level == wc->active_node)
1292                                 leave_shared_node(root, wc, *level);
1293                         *level = i + 1;
1294                 }
1295         }
1296         return 1;
1297 }
1298
1299 static int check_root_dir(struct inode_record *rec)
1300 {
1301         struct inode_backref *backref;
1302         int ret = -1;
1303
1304         if (!rec->found_inode_item || rec->errors)
1305                 goto out;
1306         if (rec->nlink != 1 || rec->found_link != 0)
1307                 goto out;
1308         if (list_empty(&rec->backrefs))
1309                 goto out;
1310         backref = list_entry(rec->backrefs.next, struct inode_backref, list);
1311         if (!backref->found_inode_ref)
1312                 goto out;
1313         if (backref->index != 0 || backref->namelen != 2 ||
1314             memcmp(backref->name, "..", 2))
1315                 goto out;
1316         if (backref->found_dir_index || backref->found_dir_item)
1317                 goto out;
1318         ret = 0;
1319 out:
1320         return ret;
1321 }
1322
1323 static int check_inode_recs(struct btrfs_root *root,
1324                             struct cache_tree *inode_cache)
1325 {
1326         struct cache_extent *cache;
1327         struct ptr_node *node;
1328         struct inode_record *rec;
1329         struct inode_backref *backref;
1330         int ret;
1331         u64 error = 0;
1332         u64 root_dirid = btrfs_root_dirid(&root->root_item);
1333
1334         if (btrfs_root_refs(&root->root_item) == 0) {
1335                 if (!cache_tree_empty(inode_cache))
1336                         fprintf(stderr, "warning line %d\n", __LINE__);
1337                 return 0;
1338         }
1339
1340         rec = get_inode_rec(inode_cache, root_dirid, 0);
1341         if (rec) {
1342                 ret = check_root_dir(rec);
1343                 if (ret) {
1344                         fprintf(stderr, "root %llu root dir %llu error\n",
1345                                 (unsigned long long)root->root_key.objectid,
1346                                 (unsigned long long)root_dirid);
1347                         error++;
1348                 }
1349         } else {
1350                 fprintf(stderr, "root %llu root dir %llu not found\n",
1351                         (unsigned long long)root->root_key.objectid,
1352                         (unsigned long long)root_dirid);
1353         }
1354
1355         while (1) {
1356                 cache = find_first_cache_extent(inode_cache, 0);
1357                 if (!cache)
1358                         break;
1359                 node = container_of(cache, struct ptr_node, cache);
1360                 rec = node->data;
1361                 remove_cache_extent(inode_cache, &node->cache);
1362                 free(node);
1363                 if (rec->ino == root_dirid ||
1364                     rec->ino == BTRFS_ORPHAN_OBJECTID) {
1365                         free_inode_rec(rec);
1366                         continue;
1367                 }
1368
1369                 if (rec->errors & I_ERR_NO_ORPHAN_ITEM) {
1370                         ret = check_orphan_item(root, rec->ino);
1371                         if (ret == 0)
1372                                 rec->errors &= ~I_ERR_NO_ORPHAN_ITEM;
1373                         if (can_free_inode_rec(rec)) {
1374                                 free_inode_rec(rec);
1375                                 continue;
1376                         }
1377                 }
1378
1379                 error++;
1380                 if (!rec->found_inode_item)
1381                         rec->errors |= I_ERR_NO_INODE_ITEM;
1382                 if (rec->found_link != rec->nlink)
1383                         rec->errors |= I_ERR_LINK_COUNT_WRONG;
1384                 fprintf(stderr, "root %llu inode %llu errors %x\n",
1385                         (unsigned long long) root->root_key.objectid,
1386                         (unsigned long long) rec->ino, rec->errors);
1387                 list_for_each_entry(backref, &rec->backrefs, list) {
1388                         if (!backref->found_dir_item)
1389                                 backref->errors |= REF_ERR_NO_DIR_ITEM;
1390                         if (!backref->found_dir_index)
1391                                 backref->errors |= REF_ERR_NO_DIR_INDEX;
1392                         if (!backref->found_inode_ref)
1393                                 backref->errors |= REF_ERR_NO_INODE_REF;
1394                         fprintf(stderr, "\tunresolved ref dir %llu index %llu"
1395                                 " namelen %u name %s filetype %d error %x\n",
1396                                 (unsigned long long)backref->dir,
1397                                 (unsigned long long)backref->index,
1398                                 backref->namelen, backref->name,
1399                                 backref->filetype, backref->errors);
1400                 }
1401                 free_inode_rec(rec);
1402         }
1403         return (error > 0) ? -1 : 0;
1404 }
1405
1406 static struct root_record *get_root_rec(struct cache_tree *root_cache,
1407                                         u64 objectid)
1408 {
1409         struct cache_extent *cache;
1410         struct root_record *rec = NULL;
1411         int ret;
1412
1413         cache = find_cache_extent(root_cache, objectid, 1);
1414         if (cache) {
1415                 rec = container_of(cache, struct root_record, cache);
1416         } else {
1417                 rec = calloc(1, sizeof(*rec));
1418                 rec->objectid = objectid;
1419                 INIT_LIST_HEAD(&rec->backrefs);
1420                 rec->cache.start = objectid;
1421                 rec->cache.size = 1;
1422
1423                 ret = insert_existing_cache_extent(root_cache, &rec->cache);
1424                 BUG_ON(ret);
1425         }
1426         return rec;
1427 }
1428
1429 static struct root_backref *get_root_backref(struct root_record *rec,
1430                                              u64 ref_root, u64 dir, u64 index,
1431                                              const char *name, int namelen)
1432 {
1433         struct root_backref *backref;
1434
1435         list_for_each_entry(backref, &rec->backrefs, list) {
1436                 if (backref->ref_root != ref_root || backref->dir != dir ||
1437                     backref->namelen != namelen)
1438                         continue;
1439                 if (memcmp(name, backref->name, namelen))
1440                         continue;
1441                 return backref;
1442         }
1443
1444         backref = malloc(sizeof(*backref) + namelen + 1);
1445         memset(backref, 0, sizeof(*backref));
1446         backref->ref_root = ref_root;
1447         backref->dir = dir;
1448         backref->index = index;
1449         backref->namelen = namelen;
1450         memcpy(backref->name, name, namelen);
1451         backref->name[namelen] = '\0';
1452         list_add_tail(&backref->list, &rec->backrefs);
1453         return backref;
1454 }
1455
1456 static void free_root_recs(struct cache_tree *root_cache)
1457 {
1458         struct cache_extent *cache;
1459         struct root_record *rec;
1460         struct root_backref *backref;
1461
1462         while (1) {
1463                 cache = find_first_cache_extent(root_cache, 0);
1464                 if (!cache)
1465                         break;
1466                 rec = container_of(cache, struct root_record, cache);
1467                 remove_cache_extent(root_cache, &rec->cache);
1468
1469                 while (!list_empty(&rec->backrefs)) {
1470                         backref = list_entry(rec->backrefs.next,
1471                                              struct root_backref, list);
1472                         list_del(&backref->list);
1473                         free(backref);
1474                 }
1475                 kfree(rec);
1476         }
1477 }
1478
1479 static int add_root_backref(struct cache_tree *root_cache,
1480                             u64 root_id, u64 ref_root, u64 dir, u64 index,
1481                             const char *name, int namelen,
1482                             int item_type, int errors)
1483 {
1484         struct root_record *rec;
1485         struct root_backref *backref;
1486
1487         rec = get_root_rec(root_cache, root_id);
1488         backref = get_root_backref(rec, ref_root, dir, index, name, namelen);
1489
1490         backref->errors |= errors;
1491
1492         if (item_type != BTRFS_DIR_ITEM_KEY) {
1493                 if (backref->found_dir_index || backref->found_back_ref ||
1494                     backref->found_forward_ref) {
1495                         if (backref->index != index)
1496                                 backref->errors |= REF_ERR_INDEX_UNMATCH;
1497                 } else {
1498                         backref->index = index;
1499                 }
1500         }
1501
1502         if (item_type == BTRFS_DIR_ITEM_KEY) {
1503                 backref->found_dir_item = 1;
1504                 backref->reachable = 1;
1505                 rec->found_ref++;
1506         } else if (item_type == BTRFS_DIR_INDEX_KEY) {
1507                 backref->found_dir_index = 1;
1508         } else if (item_type == BTRFS_ROOT_REF_KEY) {
1509                 if (backref->found_forward_ref)
1510                         backref->errors |= REF_ERR_DUP_ROOT_REF;
1511                 backref->found_forward_ref = 1;
1512         } else if (item_type == BTRFS_ROOT_BACKREF_KEY) {
1513                 if (backref->found_back_ref)
1514                         backref->errors |= REF_ERR_DUP_ROOT_BACKREF;
1515                 backref->found_back_ref = 1;
1516         } else {
1517                 BUG_ON(1);
1518         }
1519
1520         return 0;
1521 }
1522
1523 static int merge_root_recs(struct btrfs_root *root,
1524                            struct cache_tree *src_cache,
1525                            struct cache_tree *dst_cache)
1526 {
1527         struct cache_extent *cache;
1528         struct ptr_node *node;
1529         struct inode_record *rec;
1530         struct inode_backref *backref;
1531
1532         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
1533                 free_inode_recs(src_cache);
1534                 return 0;
1535         }
1536
1537         while (1) {
1538                 cache = find_first_cache_extent(src_cache, 0);
1539                 if (!cache)
1540                         break;
1541                 node = container_of(cache, struct ptr_node, cache);
1542                 rec = node->data;
1543                 remove_cache_extent(src_cache, &node->cache);
1544                 free(node);
1545
1546                 if (!is_child_root(root, root->objectid, rec->ino))
1547                         goto skip;
1548
1549                 list_for_each_entry(backref, &rec->backrefs, list) {
1550                         BUG_ON(backref->found_inode_ref);
1551                         if (backref->found_dir_item)
1552                                 add_root_backref(dst_cache, rec->ino,
1553                                         root->root_key.objectid, backref->dir,
1554                                         backref->index, backref->name,
1555                                         backref->namelen, BTRFS_DIR_ITEM_KEY,
1556                                         backref->errors);
1557                         if (backref->found_dir_index)
1558                                 add_root_backref(dst_cache, rec->ino,
1559                                         root->root_key.objectid, backref->dir,
1560                                         backref->index, backref->name,
1561                                         backref->namelen, BTRFS_DIR_INDEX_KEY,
1562                                         backref->errors);
1563                 }
1564 skip:
1565                 free_inode_rec(rec);
1566         }
1567         return 0;
1568 }
1569
1570 static int check_root_refs(struct btrfs_root *root,
1571                            struct cache_tree *root_cache)
1572 {
1573         struct root_record *rec;
1574         struct root_record *ref_root;
1575         struct root_backref *backref;
1576         struct cache_extent *cache;
1577         int loop = 1;
1578         int ret;
1579         int error;
1580         int errors = 0;
1581
1582         rec = get_root_rec(root_cache, BTRFS_FS_TREE_OBJECTID);
1583         rec->found_ref = 1;
1584
1585         /* fixme: this can not detect circular references */
1586         while (loop) {
1587                 loop = 0;
1588                 cache = find_first_cache_extent(root_cache, 0);
1589                 while (1) {
1590                         if (!cache)
1591                                 break;
1592                         rec = container_of(cache, struct root_record, cache);
1593                         cache = next_cache_extent(cache);
1594
1595                         if (rec->found_ref == 0)
1596                                 continue;
1597
1598                         list_for_each_entry(backref, &rec->backrefs, list) {
1599                                 if (!backref->reachable)
1600                                         continue;
1601
1602                                 ref_root = get_root_rec(root_cache,
1603                                                         backref->ref_root);
1604                                 if (ref_root->found_ref > 0)
1605                                         continue;
1606
1607                                 backref->reachable = 0;
1608                                 rec->found_ref--;
1609                                 if (rec->found_ref == 0)
1610                                         loop = 1;
1611                         }
1612                 }
1613         }
1614
1615         cache = find_first_cache_extent(root_cache, 0);
1616         while (1) {
1617                 if (!cache)
1618                         break;
1619                 rec = container_of(cache, struct root_record, cache);
1620                 cache = next_cache_extent(cache);
1621
1622                 if (rec->found_ref == 0 &&
1623                     rec->objectid >= BTRFS_FIRST_FREE_OBJECTID &&
1624                     rec->objectid <= BTRFS_LAST_FREE_OBJECTID) {
1625                         ret = check_orphan_item(root->fs_info->tree_root,
1626                                                 rec->objectid);
1627                         if (ret == 0)
1628                                 continue;
1629                         errors++;
1630                         fprintf(stderr, "fs tree %llu not referenced\n",
1631                                 (unsigned long long)rec->objectid);
1632                 }
1633
1634                 error = 0;
1635                 if (rec->found_ref > 0 && !rec->found_root_item)
1636                         error = 1;
1637                 list_for_each_entry(backref, &rec->backrefs, list) {
1638                         if (!backref->found_dir_item)
1639                                 backref->errors |= REF_ERR_NO_DIR_ITEM;
1640                         if (!backref->found_dir_index)
1641                                 backref->errors |= REF_ERR_NO_DIR_INDEX;
1642                         if (!backref->found_back_ref)
1643                                 backref->errors |= REF_ERR_NO_ROOT_BACKREF;
1644                         if (!backref->found_forward_ref)
1645                                 backref->errors |= REF_ERR_NO_ROOT_REF;
1646                         if (backref->reachable && backref->errors)
1647                                 error = 1;
1648                 }
1649                 if (!error)
1650                         continue;
1651
1652                 errors++;
1653                 fprintf(stderr, "fs tree %llu refs %u %s\n",
1654                         (unsigned long long)rec->objectid, rec->found_ref,
1655                          rec->found_root_item ? "" : "not found");
1656
1657                 list_for_each_entry(backref, &rec->backrefs, list) {
1658                         if (!backref->reachable)
1659                                 continue;
1660                         if (!backref->errors && rec->found_root_item)
1661                                 continue;
1662                         fprintf(stderr, "\tunresolved ref root %llu dir %llu"
1663                                 " index %llu namelen %u name %s error %x\n",
1664                                 (unsigned long long)backref->ref_root,
1665                                 (unsigned long long)backref->dir,
1666                                 (unsigned long long)backref->index,
1667                                 backref->namelen, backref->name,
1668                                 backref->errors);
1669                 }
1670         }
1671         return errors > 0 ? 1 : 0;
1672 }
1673
1674 static int process_root_ref(struct extent_buffer *eb, int slot,
1675                             struct btrfs_key *key,
1676                             struct cache_tree *root_cache)
1677 {
1678         u64 dirid;
1679         u64 index;
1680         u32 len;
1681         u32 name_len;
1682         struct btrfs_root_ref *ref;
1683         char namebuf[BTRFS_NAME_LEN];
1684         int error;
1685
1686         ref = btrfs_item_ptr(eb, slot, struct btrfs_root_ref);
1687
1688         dirid = btrfs_root_ref_dirid(eb, ref);
1689         index = btrfs_root_ref_sequence(eb, ref);
1690         name_len = btrfs_root_ref_name_len(eb, ref);
1691
1692         if (name_len <= BTRFS_NAME_LEN) {
1693                 len = name_len;
1694                 error = 0;
1695         } else {
1696                 len = BTRFS_NAME_LEN;
1697                 error = REF_ERR_NAME_TOO_LONG;
1698         }
1699         read_extent_buffer(eb, namebuf, (unsigned long)(ref + 1), len);
1700
1701         if (key->type == BTRFS_ROOT_REF_KEY) {
1702                 add_root_backref(root_cache, key->offset, key->objectid, dirid,
1703                                  index, namebuf, len, key->type, error);
1704         } else {
1705                 add_root_backref(root_cache, key->objectid, key->offset, dirid,
1706                                  index, namebuf, len, key->type, error);
1707         }
1708         return 0;
1709 }
1710
1711 static int check_fs_root(struct btrfs_root *root,
1712                          struct cache_tree *root_cache,
1713                          struct walk_control *wc)
1714 {
1715         int ret = 0;
1716         int wret;
1717         int level;
1718         struct btrfs_path path;
1719         struct shared_node root_node;
1720         struct root_record *rec;
1721         struct btrfs_root_item *root_item = &root->root_item;
1722
1723         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1724                 rec = get_root_rec(root_cache, root->root_key.objectid);
1725                 if (btrfs_root_refs(root_item) > 0)
1726                         rec->found_root_item = 1;
1727         }
1728
1729         btrfs_init_path(&path);
1730         memset(&root_node, 0, sizeof(root_node));
1731         cache_tree_init(&root_node.root_cache);
1732         cache_tree_init(&root_node.inode_cache);
1733
1734         level = btrfs_header_level(root->node);
1735         memset(wc->nodes, 0, sizeof(wc->nodes));
1736         wc->nodes[level] = &root_node;
1737         wc->active_node = level;
1738         wc->root_level = level;
1739
1740         if (btrfs_root_refs(root_item) > 0 ||
1741             btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
1742                 path.nodes[level] = root->node;
1743                 extent_buffer_get(root->node);
1744                 path.slots[level] = 0;
1745         } else {
1746                 struct btrfs_key key;
1747                 struct btrfs_disk_key found_key;
1748
1749                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
1750                 level = root_item->drop_level;
1751                 path.lowest_level = level;
1752                 wret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
1753                 BUG_ON(wret < 0);
1754                 btrfs_node_key(path.nodes[level], &found_key,
1755                                 path.slots[level]);
1756                 WARN_ON(memcmp(&found_key, &root_item->drop_progress,
1757                                         sizeof(found_key)));
1758         }
1759
1760         while (1) {
1761                 wret = walk_down_tree(root, &path, wc, &level);
1762                 if (wret < 0)
1763                         ret = wret;
1764                 if (wret != 0)
1765                         break;
1766
1767                 wret = walk_up_tree(root, &path, wc, &level);
1768                 if (wret < 0)
1769                         ret = wret;
1770                 if (wret != 0)
1771                         break;
1772         }
1773         btrfs_release_path(root, &path);
1774
1775         merge_root_recs(root, &root_node.root_cache, root_cache);
1776
1777         if (root_node.current) {
1778                 root_node.current->checked = 1;
1779                 maybe_free_inode_rec(&root_node.inode_cache,
1780                                 root_node.current);
1781         }
1782
1783         ret = check_inode_recs(root, &root_node.inode_cache);
1784         return ret;
1785 }
1786
1787 static int fs_root_objectid(u64 objectid)
1788 {
1789         if (objectid == BTRFS_FS_TREE_OBJECTID ||
1790             objectid == BTRFS_TREE_RELOC_OBJECTID ||
1791             objectid == BTRFS_DATA_RELOC_TREE_OBJECTID ||
1792             (objectid >= BTRFS_FIRST_FREE_OBJECTID &&
1793              objectid <= BTRFS_LAST_FREE_OBJECTID))
1794                 return 1;
1795         return 0;
1796 }
1797
1798 static int check_fs_roots(struct btrfs_root *root,
1799                           struct cache_tree *root_cache)
1800 {
1801         struct btrfs_path path;
1802         struct btrfs_key key;
1803         struct walk_control wc;
1804         struct extent_buffer *leaf;
1805         struct btrfs_root *tmp_root;
1806         struct btrfs_root *tree_root = root->fs_info->tree_root;
1807         int ret;
1808         int err = 0;
1809
1810         memset(&wc, 0, sizeof(wc));
1811         cache_tree_init(&wc.shared);
1812         btrfs_init_path(&path);
1813
1814         key.offset = 0;
1815         key.objectid = 0;
1816         key.type = BTRFS_ROOT_ITEM_KEY;
1817         ret = btrfs_search_slot(NULL, tree_root, &key, &path, 0, 0);
1818         BUG_ON(ret < 0);
1819         while (1) {
1820                 leaf = path.nodes[0];
1821                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1822                         ret = btrfs_next_leaf(tree_root, &path);
1823                         if (ret != 0)
1824                                 break;
1825                         leaf = path.nodes[0];
1826                 }
1827                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1828                 if (key.type == BTRFS_ROOT_ITEM_KEY &&
1829                     fs_root_objectid(key.objectid)) {
1830                         tmp_root = btrfs_read_fs_root_no_cache(root->fs_info,
1831                                                                &key);
1832                         if (IS_ERR(tmp_root)) {
1833                                 err = 1;
1834                                 goto next;
1835                         }
1836                         ret = check_fs_root(tmp_root, root_cache, &wc);
1837                         if (ret)
1838                                 err = 1;
1839                         btrfs_free_fs_root(root->fs_info, tmp_root);
1840                 } else if (key.type == BTRFS_ROOT_REF_KEY ||
1841                            key.type == BTRFS_ROOT_BACKREF_KEY) {
1842                         process_root_ref(leaf, path.slots[0], &key,
1843                                          root_cache);
1844                 }
1845 next:
1846                 path.slots[0]++;
1847         }
1848         btrfs_release_path(tree_root, &path);
1849
1850         if (!cache_tree_empty(&wc.shared))
1851                 fprintf(stderr, "warning line %d\n", __LINE__);
1852
1853         return err;
1854 }
1855
1856 static int all_backpointers_checked(struct extent_record *rec, int print_errs)
1857 {
1858         struct list_head *cur = rec->backrefs.next;
1859         struct extent_backref *back;
1860         struct tree_backref *tback;
1861         struct data_backref *dback;
1862         u64 found = 0;
1863         int err = 0;
1864
1865         while(cur != &rec->backrefs) {
1866                 back = list_entry(cur, struct extent_backref, list);
1867                 cur = cur->next;
1868                 if (!back->found_extent_tree) {
1869                         err = 1;
1870                         if (!print_errs)
1871                                 goto out;
1872                         if (back->is_data) {
1873                                 dback = (struct data_backref *)back;
1874                                 fprintf(stderr, "Backref %llu %s %llu"
1875                                         " owner %llu offset %llu num_refs %lu"
1876                                         " not found in extent tree\n",
1877                                         (unsigned long long)rec->start,
1878                                         back->full_backref ?
1879                                         "parent" : "root",
1880                                         back->full_backref ?
1881                                         (unsigned long long)dback->parent:
1882                                         (unsigned long long)dback->root,
1883                                         (unsigned long long)dback->owner,
1884                                         (unsigned long long)dback->offset,
1885                                         (unsigned long)dback->num_refs);
1886                         } else {
1887                                 tback = (struct tree_backref *)back;
1888                                 fprintf(stderr, "Backref %llu parent %llu"
1889                                         " root %llu not found in extent tree\n",
1890                                         (unsigned long long)rec->start,
1891                                         (unsigned long long)tback->parent,
1892                                         (unsigned long long)tback->root);
1893                         }
1894                 }
1895                 if (!back->is_data && !back->found_ref) {
1896                         err = 1;
1897                         if (!print_errs)
1898                                 goto out;
1899                         tback = (struct tree_backref *)back;
1900                         fprintf(stderr, "Backref %llu %s %llu not referenced back %p\n",
1901                                 (unsigned long long)rec->start,
1902                                 back->full_backref ? "parent" : "root",
1903                                 back->full_backref ?
1904                                 (unsigned long long)tback->parent :
1905                                 (unsigned long long)tback->root, back);
1906                 }
1907                 if (back->is_data) {
1908                         dback = (struct data_backref *)back;
1909                         if (dback->found_ref != dback->num_refs) {
1910                                 err = 1;
1911                                 if (!print_errs)
1912                                         goto out;
1913                                 fprintf(stderr, "Incorrect local backref count"
1914                                         " on %llu %s %llu owner %llu"
1915                                         " offset %llu found %u wanted %u back %p\n",
1916                                         (unsigned long long)rec->start,
1917                                         back->full_backref ?
1918                                         "parent" : "root",
1919                                         back->full_backref ?
1920                                         (unsigned long long)dback->parent:
1921                                         (unsigned long long)dback->root,
1922                                         (unsigned long long)dback->owner,
1923                                         (unsigned long long)dback->offset,
1924                                         dback->found_ref, dback->num_refs, back);
1925                         }
1926                         if (dback->bytes != rec->nr) {
1927                                 err = 1;
1928                                 if (!print_errs)
1929                                         goto out;
1930                                 fprintf(stderr, "Backref bytes do not match "
1931                                         "extent backref, bytenr=%llu, ref "
1932                                         "bytes=%llu, backref bytes=%llu\n",
1933                                         (unsigned long long)rec->start,
1934                                         (unsigned long long)rec->nr,
1935                                         (unsigned long long)dback->bytes);
1936                         }
1937                 }
1938                 if (!back->is_data) {
1939                         found += 1;
1940                 } else {
1941                         dback = (struct data_backref *)back;
1942                         found += dback->found_ref;
1943                 }
1944         }
1945         if (found != rec->refs) {
1946                 err = 1;
1947                 if (!print_errs)
1948                         goto out;
1949                 fprintf(stderr, "Incorrect global backref count "
1950                         "on %llu found %llu wanted %llu\n",
1951                         (unsigned long long)rec->start,
1952                         (unsigned long long)found,
1953                         (unsigned long long)rec->refs);
1954         }
1955 out:
1956         return err;
1957 }
1958
1959 static int free_all_extent_backrefs(struct extent_record *rec)
1960 {
1961         struct extent_backref *back;
1962         struct list_head *cur;
1963         while (!list_empty(&rec->backrefs)) {
1964                 cur = rec->backrefs.next;
1965                 back = list_entry(cur, struct extent_backref, list);
1966                 list_del(cur);
1967                 free(back);
1968         }
1969         return 0;
1970 }
1971
1972 static int maybe_free_extent_rec(struct cache_tree *extent_cache,
1973                                  struct extent_record *rec)
1974 {
1975         if (rec->content_checked && rec->owner_ref_checked &&
1976             rec->extent_item_refs == rec->refs && rec->refs > 0 &&
1977             !all_backpointers_checked(rec, 0)) {
1978                 remove_cache_extent(extent_cache, &rec->cache);
1979                 free_all_extent_backrefs(rec);
1980                 free(rec);
1981         }
1982         return 0;
1983 }
1984
1985 static int check_owner_ref(struct btrfs_root *root,
1986                             struct extent_record *rec,
1987                             struct extent_buffer *buf)
1988 {
1989         struct extent_backref *node;
1990         struct tree_backref *back;
1991         struct btrfs_root *ref_root;
1992         struct btrfs_key key;
1993         struct btrfs_path path;
1994         struct extent_buffer *parent;
1995         int level;
1996         int found = 0;
1997         int ret;
1998
1999         list_for_each_entry(node, &rec->backrefs, list) {
2000                 if (node->is_data)
2001                         continue;
2002                 if (!node->found_ref)
2003                         continue;
2004                 if (node->full_backref)
2005                         continue;
2006                 back = (struct tree_backref *)node;
2007                 if (btrfs_header_owner(buf) == back->root)
2008                         return 0;
2009         }
2010         BUG_ON(rec->is_root);
2011
2012         /* try to find the block by search corresponding fs tree */
2013         key.objectid = btrfs_header_owner(buf);
2014         key.type = BTRFS_ROOT_ITEM_KEY;
2015         key.offset = (u64)-1;
2016
2017         ref_root = btrfs_read_fs_root(root->fs_info, &key);
2018         if (IS_ERR(ref_root))
2019                 return 1;
2020
2021         level = btrfs_header_level(buf);
2022         if (level == 0)
2023                 btrfs_item_key_to_cpu(buf, &key, 0);
2024         else
2025                 btrfs_node_key_to_cpu(buf, &key, 0);
2026
2027         btrfs_init_path(&path);
2028         path.lowest_level = level + 1;
2029         ret = btrfs_search_slot(NULL, ref_root, &key, &path, 0, 0);
2030         if (ret < 0)
2031                 return 0;
2032
2033         parent = path.nodes[level + 1];
2034         if (parent && buf->start == btrfs_node_blockptr(parent,
2035                                                         path.slots[level + 1]))
2036                 found = 1;
2037
2038         btrfs_release_path(ref_root, &path);
2039         return found ? 0 : 1;
2040 }
2041
2042 static int is_extent_tree_record(struct extent_record *rec)
2043 {
2044         struct list_head *cur = rec->backrefs.next;
2045         struct extent_backref *node;
2046         struct tree_backref *back;
2047         int is_extent = 0;
2048
2049         while(cur != &rec->backrefs) {
2050                 node = list_entry(cur, struct extent_backref, list);
2051                 cur = cur->next;
2052                 if (node->is_data)
2053                         return 0;
2054                 back = (struct tree_backref *)node;
2055                 if (node->full_backref)
2056                         return 0;
2057                 if (back->root == BTRFS_EXTENT_TREE_OBJECTID)
2058                         is_extent = 1;
2059         }
2060         return is_extent;
2061 }
2062
2063
2064 static int record_bad_block_io(struct btrfs_fs_info *info,
2065                                struct cache_tree *extent_cache,
2066                                u64 start, u64 len)
2067 {
2068         struct extent_record *rec;
2069         struct cache_extent *cache;
2070         struct btrfs_key key;
2071
2072         cache = find_cache_extent(extent_cache, start, len);
2073         if (!cache)
2074                 return 0;
2075
2076         rec = container_of(cache, struct extent_record, cache);
2077         if (!is_extent_tree_record(rec))
2078                 return 0;
2079
2080         btrfs_disk_key_to_cpu(&key, &rec->parent_key);
2081         return btrfs_add_corrupt_extent_record(info, &key, start, len, 0);
2082 }
2083
2084 static int check_block(struct btrfs_root *root,
2085                        struct cache_tree *extent_cache,
2086                        struct extent_buffer *buf, u64 flags)
2087 {
2088         struct extent_record *rec;
2089         struct cache_extent *cache;
2090         struct btrfs_key key;
2091         int ret = 1;
2092         int level;
2093
2094         cache = find_cache_extent(extent_cache, buf->start, buf->len);
2095         if (!cache)
2096                 return 1;
2097         rec = container_of(cache, struct extent_record, cache);
2098         rec->generation = btrfs_header_generation(buf);
2099
2100         level = btrfs_header_level(buf);
2101         if (btrfs_header_nritems(buf) > 0) {
2102
2103                 if (level == 0)
2104                         btrfs_item_key_to_cpu(buf, &key, 0);
2105                 else
2106                         btrfs_node_key_to_cpu(buf, &key, 0);
2107
2108                 rec->info_objectid = key.objectid;
2109         }
2110         rec->info_level = level;
2111
2112         if (btrfs_is_leaf(buf))
2113                 ret = btrfs_check_leaf(root, &rec->parent_key, buf);
2114         else
2115                 ret = btrfs_check_node(root, &rec->parent_key, buf);
2116
2117         if (ret) {
2118                 fprintf(stderr, "bad block %llu\n",
2119                         (unsigned long long)buf->start);
2120         } else {
2121                 rec->content_checked = 1;
2122                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)
2123                         rec->owner_ref_checked = 1;
2124                 else {
2125                         ret = check_owner_ref(root, rec, buf);
2126                         if (!ret)
2127                                 rec->owner_ref_checked = 1;
2128                 }
2129         }
2130         if (!ret)
2131                 maybe_free_extent_rec(extent_cache, rec);
2132         return ret;
2133 }
2134
2135 static struct tree_backref *find_tree_backref(struct extent_record *rec,
2136                                                 u64 parent, u64 root)
2137 {
2138         struct list_head *cur = rec->backrefs.next;
2139         struct extent_backref *node;
2140         struct tree_backref *back;
2141
2142         while(cur != &rec->backrefs) {
2143                 node = list_entry(cur, struct extent_backref, list);
2144                 cur = cur->next;
2145                 if (node->is_data)
2146                         continue;
2147                 back = (struct tree_backref *)node;
2148                 if (parent > 0) {
2149                         if (!node->full_backref)
2150                                 continue;
2151                         if (parent == back->parent)
2152                                 return back;
2153                 } else {
2154                         if (node->full_backref)
2155                                 continue;
2156                         if (back->root == root)
2157                                 return back;
2158                 }
2159         }
2160         return NULL;
2161 }
2162
2163 static struct tree_backref *alloc_tree_backref(struct extent_record *rec,
2164                                                 u64 parent, u64 root)
2165 {
2166         struct tree_backref *ref = malloc(sizeof(*ref));
2167         memset(&ref->node, 0, sizeof(ref->node));
2168         if (parent > 0) {
2169                 ref->parent = parent;
2170                 ref->node.full_backref = 1;
2171         } else {
2172                 ref->root = root;
2173                 ref->node.full_backref = 0;
2174         }
2175         list_add_tail(&ref->node.list, &rec->backrefs);
2176
2177         return ref;
2178 }
2179
2180 static struct data_backref *find_data_backref(struct extent_record *rec,
2181                                                 u64 parent, u64 root,
2182                                                 u64 owner, u64 offset,
2183                                                 int found_ref, u64 bytes)
2184 {
2185         struct list_head *cur = rec->backrefs.next;
2186         struct extent_backref *node;
2187         struct data_backref *back;
2188
2189         while(cur != &rec->backrefs) {
2190                 node = list_entry(cur, struct extent_backref, list);
2191                 cur = cur->next;
2192                 if (!node->is_data)
2193                         continue;
2194                 back = (struct data_backref *)node;
2195                 if (parent > 0) {
2196                         if (!node->full_backref)
2197                                 continue;
2198                         if (parent == back->parent)
2199                                 return back;
2200                 } else {
2201                         if (node->full_backref)
2202                                 continue;
2203                         if (back->root == root && back->owner == owner &&
2204                             back->offset == offset) {
2205                                 if (found_ref && node->found_ref &&
2206                                     back->bytes != bytes)
2207                                         continue;
2208                                 return back;
2209                         }
2210                 }
2211         }
2212         return NULL;
2213 }
2214
2215 static struct data_backref *alloc_data_backref(struct extent_record *rec,
2216                                                 u64 parent, u64 root,
2217                                                 u64 owner, u64 offset,
2218                                                 u64 max_size)
2219 {
2220         struct data_backref *ref = malloc(sizeof(*ref));
2221         memset(&ref->node, 0, sizeof(ref->node));
2222         ref->node.is_data = 1;
2223
2224         if (parent > 0) {
2225                 ref->parent = parent;
2226                 ref->owner = 0;
2227                 ref->offset = 0;
2228                 ref->node.full_backref = 1;
2229         } else {
2230                 ref->root = root;
2231                 ref->owner = owner;
2232                 ref->offset = offset;
2233                 ref->node.full_backref = 0;
2234         }
2235         ref->bytes = max_size;
2236         ref->found_ref = 0;
2237         ref->num_refs = 0;
2238         list_add_tail(&ref->node.list, &rec->backrefs);
2239         if (max_size > rec->max_size)
2240                 rec->max_size = max_size;
2241         return ref;
2242 }
2243
2244 static int add_extent_rec(struct cache_tree *extent_cache,
2245                           struct btrfs_key *parent_key,
2246                           u64 start, u64 nr, u64 extent_item_refs,
2247                           int is_root, int inc_ref, int set_checked,
2248                           int metadata, int extent_rec, u64 max_size)
2249 {
2250         struct extent_record *rec;
2251         struct cache_extent *cache;
2252         int ret = 0;
2253
2254         cache = find_cache_extent(extent_cache, start, nr);
2255         if (cache) {
2256                 rec = container_of(cache, struct extent_record, cache);
2257                 if (inc_ref)
2258                         rec->refs++;
2259                 if (rec->nr == 1)
2260                         rec->nr = max(nr, max_size);
2261
2262                 /*
2263                  * We need to make sure to reset nr to whatever the extent
2264                  * record says was the real size, this way we can compare it to
2265                  * the backrefs.
2266                  */
2267                 if (extent_rec)
2268                         rec->nr = nr;
2269
2270                 if (start != rec->start) {
2271                         fprintf(stderr, "warning, start mismatch %llu %llu\n",
2272                                 (unsigned long long)rec->start,
2273                                 (unsigned long long)start);
2274                         ret = 1;
2275                 }
2276                 if (extent_item_refs) {
2277                         if (rec->extent_item_refs) {
2278                                 fprintf(stderr, "block %llu rec "
2279                                         "extent_item_refs %llu, passed %llu\n",
2280                                         (unsigned long long)start,
2281                                         (unsigned long long)
2282                                                         rec->extent_item_refs,
2283                                         (unsigned long long)extent_item_refs);
2284                         }
2285                         rec->extent_item_refs = extent_item_refs;
2286                 }
2287                 if (is_root)
2288                         rec->is_root = 1;
2289                 if (set_checked) {
2290                         rec->content_checked = 1;
2291                         rec->owner_ref_checked = 1;
2292                 }
2293
2294                 if (parent_key)
2295                         btrfs_cpu_key_to_disk(&rec->parent_key, parent_key);
2296
2297                 if (rec->max_size < max_size)
2298                         rec->max_size = max_size;
2299
2300                 maybe_free_extent_rec(extent_cache, rec);
2301                 return ret;
2302         }
2303         rec = malloc(sizeof(*rec));
2304         rec->start = start;
2305         rec->max_size = max_size;
2306         rec->nr = max(nr, max_size);
2307         rec->content_checked = 0;
2308         rec->owner_ref_checked = 0;
2309         rec->metadata = metadata;
2310         INIT_LIST_HEAD(&rec->backrefs);
2311
2312         if (is_root)
2313                 rec->is_root = 1;
2314         else
2315                 rec->is_root = 0;
2316
2317         if (inc_ref)
2318                 rec->refs = 1;
2319         else
2320                 rec->refs = 0;
2321
2322         if (extent_item_refs)
2323                 rec->extent_item_refs = extent_item_refs;
2324         else
2325                 rec->extent_item_refs = 0;
2326
2327         if (parent_key)
2328                 btrfs_cpu_key_to_disk(&rec->parent_key, parent_key);
2329         else
2330                 memset(&rec->parent_key, 0, sizeof(*parent_key));
2331
2332         rec->cache.start = start;
2333         rec->cache.size = nr;
2334         ret = insert_existing_cache_extent(extent_cache, &rec->cache);
2335         BUG_ON(ret);
2336         bytes_used += nr;
2337         if (set_checked) {
2338                 rec->content_checked = 1;
2339                 rec->owner_ref_checked = 1;
2340         }
2341         return ret;
2342 }
2343
2344 static int add_tree_backref(struct cache_tree *extent_cache, u64 bytenr,
2345                             u64 parent, u64 root, int found_ref)
2346 {
2347         struct extent_record *rec;
2348         struct tree_backref *back;
2349         struct cache_extent *cache;
2350
2351         cache = find_cache_extent(extent_cache, bytenr, 1);
2352         if (!cache) {
2353                 add_extent_rec(extent_cache, NULL, bytenr,
2354                                1, 0, 0, 0, 0, 1, 0, 0);
2355                 cache = find_cache_extent(extent_cache, bytenr, 1);
2356                 if (!cache)
2357                         abort();
2358         }
2359
2360         rec = container_of(cache, struct extent_record, cache);
2361         if (rec->start != bytenr) {
2362                 abort();
2363         }
2364
2365         back = find_tree_backref(rec, parent, root);
2366         if (!back)
2367                 back = alloc_tree_backref(rec, parent, root);
2368
2369         if (found_ref) {
2370                 if (back->node.found_ref) {
2371                         fprintf(stderr, "Extent back ref already exists "
2372                                 "for %llu parent %llu root %llu \n",
2373                                 (unsigned long long)bytenr,
2374                                 (unsigned long long)parent,
2375                                 (unsigned long long)root);
2376                 }
2377                 back->node.found_ref = 1;
2378         } else {
2379                 if (back->node.found_extent_tree) {
2380                         fprintf(stderr, "Extent back ref already exists "
2381                                 "for %llu parent %llu root %llu \n",
2382                                 (unsigned long long)bytenr,
2383                                 (unsigned long long)parent,
2384                                 (unsigned long long)root);
2385                 }
2386                 back->node.found_extent_tree = 1;
2387         }
2388         return 0;
2389 }
2390
2391 static int add_data_backref(struct cache_tree *extent_cache, u64 bytenr,
2392                             u64 parent, u64 root, u64 owner, u64 offset,
2393                             u32 num_refs, int found_ref, u64 max_size)
2394 {
2395         struct extent_record *rec;
2396         struct data_backref *back;
2397         struct cache_extent *cache;
2398
2399         cache = find_cache_extent(extent_cache, bytenr, 1);
2400         if (!cache) {
2401                 add_extent_rec(extent_cache, NULL, bytenr, 1, 0, 0, 0, 0,
2402                                0, 0, max_size);
2403                 cache = find_cache_extent(extent_cache, bytenr, 1);
2404                 if (!cache)
2405                         abort();
2406         }
2407
2408         rec = container_of(cache, struct extent_record, cache);
2409         if (rec->start != bytenr) {
2410                 abort();
2411         }
2412         if (rec->max_size < max_size)
2413                 rec->max_size = max_size;
2414
2415         /*
2416          * If found_ref is set then max_size is the real size and must match the
2417          * existing refs.  So if we have already found a ref then we need to
2418          * make sure that this ref matches the existing one, otherwise we need
2419          * to add a new backref so we can notice that the backrefs don't match
2420          * and we need to figure out who is telling the truth.  This is to
2421          * account for that awful fsync bug I introduced where we'd end up with
2422          * a btrfs_file_extent_item that would have its length include multiple
2423          * prealloc extents or point inside of a prealloc extent.
2424          */
2425         back = find_data_backref(rec, parent, root, owner, offset, found_ref,
2426                                  max_size);
2427         if (!back)
2428                 back = alloc_data_backref(rec, parent, root, owner, offset,
2429                                           max_size);
2430
2431         if (found_ref) {
2432                 BUG_ON(num_refs != 1);
2433                 if (back->node.found_ref)
2434                         BUG_ON(back->bytes != max_size);
2435                 back->node.found_ref = 1;
2436                 back->found_ref += 1;
2437                 back->bytes = max_size;
2438         } else {
2439                 if (back->node.found_extent_tree) {
2440                         fprintf(stderr, "Extent back ref already exists "
2441                                 "for %llu parent %llu root %llu"
2442                                 "owner %llu offset %llu num_refs %lu\n",
2443                                 (unsigned long long)bytenr,
2444                                 (unsigned long long)parent,
2445                                 (unsigned long long)root,
2446                                 (unsigned long long)owner,
2447                                 (unsigned long long)offset,
2448                                 (unsigned long)num_refs);
2449                 }
2450                 back->num_refs = num_refs;
2451                 back->node.found_extent_tree = 1;
2452         }
2453         return 0;
2454 }
2455
2456 static int add_pending(struct cache_tree *pending,
2457                        struct cache_tree *seen, u64 bytenr, u32 size)
2458 {
2459         int ret;
2460         ret = insert_cache_extent(seen, bytenr, size);
2461         if (ret)
2462                 return ret;
2463         insert_cache_extent(pending, bytenr, size);
2464         return 0;
2465 }
2466
2467 static int pick_next_pending(struct cache_tree *pending,
2468                         struct cache_tree *reada,
2469                         struct cache_tree *nodes,
2470                         u64 last, struct block_info *bits, int bits_nr,
2471                         int *reada_bits)
2472 {
2473         unsigned long node_start = last;
2474         struct cache_extent *cache;
2475         int ret;
2476
2477         cache = find_first_cache_extent(reada, 0);
2478         if (cache) {
2479                 bits[0].start = cache->start;
2480                 bits[1].size = cache->size;
2481                 *reada_bits = 1;
2482                 return 1;
2483         }
2484         *reada_bits = 0;
2485         if (node_start > 32768)
2486                 node_start -= 32768;
2487
2488         cache = find_first_cache_extent(nodes, node_start);
2489         if (!cache)
2490                 cache = find_first_cache_extent(nodes, 0);
2491
2492         if (!cache) {
2493                  cache = find_first_cache_extent(pending, 0);
2494                  if (!cache)
2495                          return 0;
2496                  ret = 0;
2497                  do {
2498                          bits[ret].start = cache->start;
2499                          bits[ret].size = cache->size;
2500                          cache = next_cache_extent(cache);
2501                          ret++;
2502                  } while (cache && ret < bits_nr);
2503                  return ret;
2504         }
2505
2506         ret = 0;
2507         do {
2508                 bits[ret].start = cache->start;
2509                 bits[ret].size = cache->size;
2510                 cache = next_cache_extent(cache);
2511                 ret++;
2512         } while (cache && ret < bits_nr);
2513
2514         if (bits_nr - ret > 8) {
2515                 u64 lookup = bits[0].start + bits[0].size;
2516                 struct cache_extent *next;
2517                 next = find_first_cache_extent(pending, lookup);
2518                 while(next) {
2519                         if (next->start - lookup > 32768)
2520                                 break;
2521                         bits[ret].start = next->start;
2522                         bits[ret].size = next->size;
2523                         lookup = next->start + next->size;
2524                         ret++;
2525                         if (ret == bits_nr)
2526                                 break;
2527                         next = next_cache_extent(next);
2528                         if (!next)
2529                                 break;
2530                 }
2531         }
2532         return ret;
2533 }
2534
2535 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2536 static int process_extent_ref_v0(struct cache_tree *extent_cache,
2537                                  struct extent_buffer *leaf, int slot)
2538 {
2539         struct btrfs_extent_ref_v0 *ref0;
2540         struct btrfs_key key;
2541
2542         btrfs_item_key_to_cpu(leaf, &key, slot);
2543         ref0 = btrfs_item_ptr(leaf, slot, struct btrfs_extent_ref_v0);
2544         if (btrfs_ref_objectid_v0(leaf, ref0) < BTRFS_FIRST_FREE_OBJECTID) {
2545                 add_tree_backref(extent_cache, key.objectid, key.offset, 0, 0);
2546         } else {
2547                 add_data_backref(extent_cache, key.objectid, key.offset, 0,
2548                                  0, 0, btrfs_ref_count_v0(leaf, ref0), 0, 0);
2549         }
2550         return 0;
2551 }
2552 #endif
2553
2554 static int process_extent_item(struct btrfs_root *root,
2555                                struct cache_tree *extent_cache,
2556                                struct extent_buffer *eb, int slot)
2557 {
2558         struct btrfs_extent_item *ei;
2559         struct btrfs_extent_inline_ref *iref;
2560         struct btrfs_extent_data_ref *dref;
2561         struct btrfs_shared_data_ref *sref;
2562         struct btrfs_key key;
2563         unsigned long end;
2564         unsigned long ptr;
2565         int type;
2566         u32 item_size = btrfs_item_size_nr(eb, slot);
2567         u64 refs = 0;
2568         u64 offset;
2569         u64 num_bytes;
2570         int metadata = 0;
2571
2572         btrfs_item_key_to_cpu(eb, &key, slot);
2573
2574         if (key.type == BTRFS_METADATA_ITEM_KEY) {
2575                 metadata = 1;
2576                 num_bytes = root->leafsize;
2577         } else {
2578                 num_bytes = key.offset;
2579         }
2580
2581         if (item_size < sizeof(*ei)) {
2582 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2583                 struct btrfs_extent_item_v0 *ei0;
2584                 BUG_ON(item_size != sizeof(*ei0));
2585                 ei0 = btrfs_item_ptr(eb, slot, struct btrfs_extent_item_v0);
2586                 refs = btrfs_extent_refs_v0(eb, ei0);
2587 #else
2588                 BUG();
2589 #endif
2590                 return add_extent_rec(extent_cache, NULL, key.objectid,
2591                                       num_bytes, refs, 0, 0, 0, metadata, 1,
2592                                       num_bytes);
2593         }
2594
2595         ei = btrfs_item_ptr(eb, slot, struct btrfs_extent_item);
2596         refs = btrfs_extent_refs(eb, ei);
2597
2598         add_extent_rec(extent_cache, NULL, key.objectid, num_bytes,
2599                        refs, 0, 0, 0, metadata, 1, num_bytes);
2600
2601         ptr = (unsigned long)(ei + 1);
2602         if (btrfs_extent_flags(eb, ei) & BTRFS_EXTENT_FLAG_TREE_BLOCK &&
2603             key.type == BTRFS_EXTENT_ITEM_KEY)
2604                 ptr += sizeof(struct btrfs_tree_block_info);
2605
2606         end = (unsigned long)ei + item_size;
2607         while (ptr < end) {
2608                 iref = (struct btrfs_extent_inline_ref *)ptr;
2609                 type = btrfs_extent_inline_ref_type(eb, iref);
2610                 offset = btrfs_extent_inline_ref_offset(eb, iref);
2611                 switch (type) {
2612                 case BTRFS_TREE_BLOCK_REF_KEY:
2613                         add_tree_backref(extent_cache, key.objectid,
2614                                          0, offset, 0);
2615                         break;
2616                 case BTRFS_SHARED_BLOCK_REF_KEY:
2617                         add_tree_backref(extent_cache, key.objectid,
2618                                          offset, 0, 0);
2619                         break;
2620                 case BTRFS_EXTENT_DATA_REF_KEY:
2621                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2622                         add_data_backref(extent_cache, key.objectid, 0,
2623                                         btrfs_extent_data_ref_root(eb, dref),
2624                                         btrfs_extent_data_ref_objectid(eb,
2625                                                                        dref),
2626                                         btrfs_extent_data_ref_offset(eb, dref),
2627                                         btrfs_extent_data_ref_count(eb, dref),
2628                                         0, num_bytes);
2629                         break;
2630                 case BTRFS_SHARED_DATA_REF_KEY:
2631                         sref = (struct btrfs_shared_data_ref *)(iref + 1);
2632                         add_data_backref(extent_cache, key.objectid, offset,
2633                                         0, 0, 0,
2634                                         btrfs_shared_data_ref_count(eb, sref),
2635                                         0, num_bytes);
2636                         break;
2637                 default:
2638                         fprintf(stderr, "corrupt extent record: key %Lu %u %Lu\n",
2639                                 key.objectid, key.type, num_bytes);
2640                         goto out;
2641                 }
2642                 ptr += btrfs_extent_inline_ref_size(type);
2643         }
2644         WARN_ON(ptr > end);
2645 out:
2646         return 0;
2647 }
2648
2649 static int check_cache_range(struct btrfs_root *root,
2650                              struct btrfs_block_group_cache *cache,
2651                              u64 offset, u64 bytes)
2652 {
2653         struct btrfs_free_space *entry;
2654         u64 *logical;
2655         u64 bytenr;
2656         int stripe_len;
2657         int i, nr, ret;
2658
2659         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2660                 bytenr = btrfs_sb_offset(i);
2661                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
2662                                        cache->key.objectid, bytenr, 0,
2663                                        &logical, &nr, &stripe_len);
2664                 if (ret)
2665                         return ret;
2666
2667                 while (nr--) {
2668                         if (logical[nr] + stripe_len <= offset)
2669                                 continue;
2670                         if (offset + bytes <= logical[nr])
2671                                 continue;
2672                         if (logical[nr] == offset) {
2673                                 if (stripe_len >= bytes) {
2674                                         kfree(logical);
2675                                         return 0;
2676                                 }
2677                                 bytes -= stripe_len;
2678                                 offset += stripe_len;
2679                         } else if (logical[nr] < offset) {
2680                                 if (logical[nr] + stripe_len >=
2681                                     offset + bytes) {
2682                                         kfree(logical);
2683                                         return 0;
2684                                 }
2685                                 bytes = (offset + bytes) -
2686                                         (logical[nr] + stripe_len);
2687                                 offset = logical[nr] + stripe_len;
2688                         } else {
2689                                 /*
2690                                  * Could be tricky, the super may land in the
2691                                  * middle of the area we're checking.  First
2692                                  * check the easiest case, it's at the end.
2693                                  */
2694                                 if (logical[nr] + stripe_len >=
2695                                     bytes + offset) {
2696                                         bytes = logical[nr] - offset;
2697                                         continue;
2698                                 }
2699
2700                                 /* Check the left side */
2701                                 ret = check_cache_range(root, cache,
2702                                                         offset,
2703                                                         logical[nr] - offset);
2704                                 if (ret) {
2705                                         kfree(logical);
2706                                         return ret;
2707                                 }
2708
2709                                 /* Now we continue with the right side */
2710                                 bytes = (offset + bytes) -
2711                                         (logical[nr] + stripe_len);
2712                                 offset = logical[nr] + stripe_len;
2713                         }
2714                 }
2715
2716                 kfree(logical);
2717         }
2718
2719         entry = btrfs_find_free_space(cache->free_space_ctl, offset, bytes);
2720         if (!entry) {
2721                 fprintf(stderr, "There is no free space entry for %Lu-%Lu\n",
2722                         offset, offset+bytes);
2723                 return -EINVAL;
2724         }
2725
2726         if (entry->offset != offset) {
2727                 fprintf(stderr, "Wanted offset %Lu, found %Lu\n", offset,
2728                         entry->offset);
2729                 return -EINVAL;
2730         }
2731
2732         if (entry->bytes != bytes) {
2733                 fprintf(stderr, "Wanted bytes %Lu, found %Lu for off %Lu\n",
2734                         bytes, entry->bytes, offset);
2735                 return -EINVAL;
2736         }
2737
2738         unlink_free_space(cache->free_space_ctl, entry);
2739         free(entry);
2740         return 0;
2741 }
2742
2743 static int verify_space_cache(struct btrfs_root *root,
2744                               struct btrfs_block_group_cache *cache)
2745 {
2746         struct btrfs_path *path;
2747         struct extent_buffer *leaf;
2748         struct btrfs_key key;
2749         u64 last;
2750         int ret = 0;
2751
2752         path = btrfs_alloc_path();
2753         if (!path)
2754                 return -ENOMEM;
2755
2756         root = root->fs_info->extent_root;
2757
2758         last = max_t(u64, cache->key.objectid, BTRFS_SUPER_INFO_OFFSET);
2759
2760         key.objectid = last;
2761         key.offset = 0;
2762         key.type = BTRFS_EXTENT_ITEM_KEY;
2763
2764         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2765         if (ret < 0)
2766                 return ret;
2767         ret = 0;
2768         while (1) {
2769                 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2770                         ret = btrfs_next_leaf(root, path);
2771                         if (ret < 0)
2772                                 return ret;
2773                         if (ret > 0) {
2774                                 ret = 0;
2775                                 break;
2776                         }
2777                 }
2778                 leaf = path->nodes[0];
2779                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2780                 if (key.objectid >= cache->key.offset + cache->key.objectid)
2781                         break;
2782                 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2783                     key.type != BTRFS_METADATA_ITEM_KEY) {
2784                         path->slots[0]++;
2785                         continue;
2786                 }
2787
2788                 if (last == key.objectid) {
2789                         last = key.objectid + key.offset;
2790                         path->slots[0]++;
2791                         continue;
2792                 }
2793
2794                 ret = check_cache_range(root, cache, last,
2795                                         key.objectid - last);
2796                 if (ret)
2797                         break;
2798                 if (key.type == BTRFS_EXTENT_ITEM_KEY)
2799                         last = key.objectid + key.offset;
2800                 else
2801                         last = key.objectid + root->leafsize;
2802                 path->slots[0]++;
2803         }
2804
2805         if (last < cache->key.objectid + cache->key.offset)
2806                 ret = check_cache_range(root, cache, last,
2807                                         cache->key.objectid +
2808                                         cache->key.offset - last);
2809         btrfs_free_path(path);
2810
2811         if (!ret &&
2812             !RB_EMPTY_ROOT(&cache->free_space_ctl->free_space_offset)) {
2813                 fprintf(stderr, "There are still entries left in the space "
2814                         "cache\n");
2815                 ret = -EINVAL;
2816         }
2817
2818         return ret;
2819 }
2820
2821 static int check_space_cache(struct btrfs_root *root)
2822 {
2823         struct btrfs_block_group_cache *cache;
2824         u64 start = BTRFS_SUPER_INFO_OFFSET + BTRFS_SUPER_INFO_SIZE;
2825         int ret;
2826         int error = 0;
2827
2828         if (btrfs_super_generation(root->fs_info->super_copy) !=
2829             btrfs_super_cache_generation(root->fs_info->super_copy)) {
2830                 printf("cache and super generation don't match, space cache "
2831                        "will be invalidated\n");
2832                 return 0;
2833         }
2834
2835         while (1) {
2836                 cache = btrfs_lookup_first_block_group(root->fs_info, start);
2837                 if (!cache)
2838                         break;
2839
2840                 start = cache->key.objectid + cache->key.offset;
2841                 if (!cache->free_space_ctl) {
2842                         int sectorsize;
2843
2844                         if (cache->flags & (BTRFS_BLOCK_GROUP_METADATA |
2845                                             BTRFS_BLOCK_GROUP_SYSTEM))
2846                                 sectorsize = root->leafsize;
2847                         else
2848                                 sectorsize = root->sectorsize;
2849
2850                         if (btrfs_init_free_space_ctl(cache, sectorsize)) {
2851                                 ret = -ENOMEM;
2852                                 break;
2853                         }
2854                 } else {
2855                         btrfs_remove_free_space_cache(cache);
2856                 }
2857
2858                 ret = load_free_space_cache(root->fs_info, cache);
2859                 if (!ret)
2860                         continue;
2861
2862                 ret = verify_space_cache(root, cache);
2863                 if (ret) {
2864                         fprintf(stderr, "cache appears valid but isnt %Lu\n",
2865                                 cache->key.objectid);
2866                         error++;
2867                 }
2868         }
2869
2870         return error ? -EINVAL : 0;
2871 }
2872
2873 static int check_extent_exists(struct btrfs_root *root, u64 bytenr,
2874                                u64 num_bytes)
2875 {
2876         struct btrfs_path *path;
2877         struct extent_buffer *leaf;
2878         struct btrfs_key key;
2879         int ret;
2880
2881         path = btrfs_alloc_path();
2882         if (!path) {
2883                 fprintf(stderr, "Error allocing path\n");
2884                 return -ENOMEM;
2885         }
2886
2887         key.objectid = bytenr;
2888         key.type = BTRFS_EXTENT_ITEM_KEY;
2889         key.offset = 0;
2890
2891
2892 again:
2893         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
2894                                 0, 0);
2895         if (ret < 0) {
2896                 fprintf(stderr, "Error looking up extent record %d\n", ret);
2897                 btrfs_free_path(path);
2898                 return ret;
2899         } else if (ret) {
2900                 if (path->slots[0])
2901                         path->slots[0]--;
2902                 else
2903                         btrfs_prev_leaf(root, path);
2904         }
2905
2906         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2907
2908         /*
2909          * Block group items come before extent items if they have the same
2910          * bytenr, so walk back one more just in case.  Dear future traveler,
2911          * first congrats on mastering time travel.  Now if it's not too much
2912          * trouble could you go back to 2006 and tell Chris to make the
2913          * BLOCK_GROUP_ITEM_KEY lower than the EXTENT_ITEM_KEY please?
2914          */
2915         if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2916                 if (path->slots[0])
2917                         path->slots[0]--;
2918                 else
2919                         btrfs_prev_leaf(root, path);
2920         }
2921
2922         while (num_bytes) {
2923                 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2924                         ret = btrfs_next_leaf(root, path);
2925                         if (ret < 0) {
2926                                 fprintf(stderr, "Error going to next leaf "
2927                                         "%d\n", ret);
2928                                 btrfs_free_path(path);
2929                                 return ret;
2930                         } else if (ret) {
2931                                 break;
2932                         }
2933                 }
2934                 leaf = path->nodes[0];
2935                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2936                 if (key.type != BTRFS_EXTENT_ITEM_KEY) {
2937                         path->slots[0]++;
2938                         continue;
2939                 }
2940                 if (key.objectid + key.offset < bytenr) {
2941                         path->slots[0]++;
2942                         continue;
2943                 }
2944                 if (key.objectid > bytenr + num_bytes)
2945                         break;
2946
2947                 if (key.objectid == bytenr) {
2948                         if (key.offset >= num_bytes) {
2949                                 num_bytes = 0;
2950                                 break;
2951                         }
2952                         num_bytes -= key.offset;
2953                         bytenr += key.offset;
2954                 } else if (key.objectid < bytenr) {
2955                         if (key.objectid + key.offset >= bytenr + num_bytes) {
2956                                 num_bytes = 0;
2957                                 break;
2958                         }
2959                         num_bytes = (bytenr + num_bytes) -
2960                                 (key.objectid + key.offset);
2961                         bytenr = key.objectid + key.offset;
2962                 } else {
2963                         if (key.objectid + key.offset < bytenr + num_bytes) {
2964                                 u64 new_start = key.objectid + key.offset;
2965                                 u64 new_bytes = bytenr + num_bytes - new_start;
2966
2967                                 /*
2968                                  * Weird case, the extent is in the middle of
2969                                  * our range, we'll have to search one side
2970                                  * and then the other.  Not sure if this happens
2971                                  * in real life, but no harm in coding it up
2972                                  * anyway just in case.
2973                                  */
2974                                 btrfs_release_path(root, path);
2975                                 ret = check_extent_exists(root, new_start,
2976                                                           new_bytes);
2977                                 if (ret) {
2978                                         fprintf(stderr, "Right section didn't "
2979                                                 "have a record\n");
2980                                         break;
2981                                 }
2982                                 num_bytes = key.objectid - bytenr;
2983                                 goto again;
2984                         }
2985                         num_bytes = key.objectid - bytenr;
2986                 }
2987                 path->slots[0]++;
2988         }
2989         ret = 0;
2990
2991         if (num_bytes) {
2992                 fprintf(stderr, "There are no extents for csum range "
2993                         "%Lu-%Lu\n", bytenr, bytenr+num_bytes);
2994                 ret = 1;
2995         }
2996
2997         btrfs_free_path(path);
2998         return ret;
2999 }
3000
3001 static int check_csums(struct btrfs_root *root)
3002 {
3003         struct btrfs_path *path;
3004         struct extent_buffer *leaf;
3005         struct btrfs_key key;
3006         u64 offset = 0, num_bytes = 0;
3007         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
3008         int errors = 0;
3009         int ret;
3010
3011         root = root->fs_info->csum_root;
3012
3013         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3014         key.type = BTRFS_EXTENT_CSUM_KEY;
3015         key.offset = 0;
3016
3017         path = btrfs_alloc_path();
3018         if (!path)
3019                 return -ENOMEM;
3020
3021         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3022         if (ret < 0) {
3023                 fprintf(stderr, "Error searching csum tree %d\n", ret);
3024                 btrfs_free_path(path);
3025                 return ret;
3026         }
3027
3028         if (ret > 0 && path->slots[0])
3029                 path->slots[0]--;
3030         ret = 0;
3031
3032         while (1) {
3033                 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
3034                         ret = btrfs_next_leaf(root, path);
3035                         if (ret < 0) {
3036                                 fprintf(stderr, "Error going to next leaf "
3037                                         "%d\n", ret);
3038                                 break;
3039                         }
3040                         if (ret)
3041                                 break;
3042                 }
3043                 leaf = path->nodes[0];
3044
3045                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3046                 if (key.type != BTRFS_EXTENT_CSUM_KEY) {
3047                         path->slots[0]++;
3048                         continue;
3049                 }
3050
3051                 if (!num_bytes) {
3052                         offset = key.offset;
3053                 } else if (key.offset != offset + num_bytes) {
3054                         ret = check_extent_exists(root, offset, num_bytes);
3055                         if (ret) {
3056                                 fprintf(stderr, "Csum exists for %Lu-%Lu but "
3057                                         "there is no extent record\n",
3058                                         offset, offset+num_bytes);
3059                                 errors++;
3060                         }
3061                         offset = key.offset;
3062                         num_bytes = 0;
3063                 }
3064
3065                 num_bytes += (btrfs_item_size_nr(leaf, path->slots[0]) /
3066                               csum_size) * root->sectorsize;
3067                 path->slots[0]++;
3068         }
3069
3070         btrfs_free_path(path);
3071         return errors;
3072 }
3073
3074 static int run_next_block(struct btrfs_root *root,
3075                           struct block_info *bits,
3076                           int bits_nr,
3077                           u64 *last,
3078                           struct cache_tree *pending,
3079                           struct cache_tree *seen,
3080                           struct cache_tree *reada,
3081                           struct cache_tree *nodes,
3082                           struct cache_tree *extent_cache)
3083 {
3084         struct extent_buffer *buf;
3085         u64 bytenr;
3086         u32 size;
3087         u64 parent;
3088         u64 owner;
3089         u64 flags;
3090         int ret;
3091         int i;
3092         int nritems;
3093         struct btrfs_key key;
3094         struct cache_extent *cache;
3095         int reada_bits;
3096
3097         ret = pick_next_pending(pending, reada, nodes, *last, bits,
3098                                 bits_nr, &reada_bits);
3099         if (ret == 0) {
3100                 return 1;
3101         }
3102         if (!reada_bits) {
3103                 for(i = 0; i < ret; i++) {
3104                         insert_cache_extent(reada, bits[i].start,
3105                                             bits[i].size);
3106
3107                         /* fixme, get the parent transid */
3108                         readahead_tree_block(root, bits[i].start,
3109                                              bits[i].size, 0);
3110                 }
3111         }
3112         *last = bits[0].start;
3113         bytenr = bits[0].start;
3114         size = bits[0].size;
3115
3116         cache = find_cache_extent(pending, bytenr, size);
3117         if (cache) {
3118                 remove_cache_extent(pending, cache);
3119                 free(cache);
3120         }
3121         cache = find_cache_extent(reada, bytenr, size);
3122         if (cache) {
3123                 remove_cache_extent(reada, cache);
3124                 free(cache);
3125         }
3126         cache = find_cache_extent(nodes, bytenr, size);
3127         if (cache) {
3128                 remove_cache_extent(nodes, cache);
3129                 free(cache);
3130         }
3131
3132         /* fixme, get the real parent transid */
3133         buf = read_tree_block(root, bytenr, size, 0);
3134         if (!extent_buffer_uptodate(buf)) {
3135                 record_bad_block_io(root->fs_info,
3136                                     extent_cache, bytenr, size);
3137                 goto out;
3138         }
3139
3140         nritems = btrfs_header_nritems(buf);
3141
3142         ret = btrfs_lookup_extent_info(NULL, root, bytenr,
3143                                        btrfs_header_level(buf), 1, NULL,
3144                                        &flags);
3145         if (ret < 0)
3146                 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
3147
3148         if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
3149                 parent = bytenr;
3150                 owner = 0;
3151         } else {
3152                 parent = 0;
3153                 owner = btrfs_header_owner(buf);
3154         }
3155
3156         ret = check_block(root, extent_cache, buf, flags);
3157         if (ret)
3158                 goto out;
3159
3160         if (btrfs_is_leaf(buf)) {
3161                 btree_space_waste += btrfs_leaf_free_space(root, buf);
3162                 for (i = 0; i < nritems; i++) {
3163                         struct btrfs_file_extent_item *fi;
3164                         btrfs_item_key_to_cpu(buf, &key, i);
3165                         if (key.type == BTRFS_EXTENT_ITEM_KEY) {
3166                                 process_extent_item(root, extent_cache, buf,
3167                                                     i);
3168                                 continue;
3169                         }
3170                         if (key.type == BTRFS_METADATA_ITEM_KEY) {
3171                                 process_extent_item(root, extent_cache, buf,
3172                                                     i);
3173                                 continue;
3174                         }
3175                         if (key.type == BTRFS_EXTENT_CSUM_KEY) {
3176                                 total_csum_bytes +=
3177                                         btrfs_item_size_nr(buf, i);
3178                                 continue;
3179                         }
3180                         if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
3181                                 continue;
3182                         }
3183                         if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
3184 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3185                                 process_extent_ref_v0(extent_cache, buf, i);
3186 #else
3187                                 BUG();
3188 #endif
3189                                 continue;
3190                         }
3191
3192                         if (key.type == BTRFS_TREE_BLOCK_REF_KEY) {
3193                                 add_tree_backref(extent_cache, key.objectid, 0,
3194                                                  key.offset, 0);
3195                                 continue;
3196                         }
3197                         if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) {
3198                                 add_tree_backref(extent_cache, key.objectid,
3199                                                  key.offset, 0, 0);
3200                                 continue;
3201                         }
3202                         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
3203                                 struct btrfs_extent_data_ref *ref;
3204                                 ref = btrfs_item_ptr(buf, i,
3205                                                 struct btrfs_extent_data_ref);
3206                                 add_data_backref(extent_cache,
3207                                         key.objectid, 0,
3208                                         btrfs_extent_data_ref_root(buf, ref),
3209                                         btrfs_extent_data_ref_objectid(buf,
3210                                                                        ref),
3211                                         btrfs_extent_data_ref_offset(buf, ref),
3212                                         btrfs_extent_data_ref_count(buf, ref),
3213                                         0, root->sectorsize);
3214                                 continue;
3215                         }
3216                         if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
3217                                 struct btrfs_shared_data_ref *ref;
3218                                 ref = btrfs_item_ptr(buf, i,
3219                                                 struct btrfs_shared_data_ref);
3220                                 add_data_backref(extent_cache,
3221                                         key.objectid, key.offset, 0, 0, 0,
3222                                         btrfs_shared_data_ref_count(buf, ref),
3223                                         0, root->sectorsize);
3224                                 continue;
3225                         }
3226                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3227                                 continue;
3228                         fi = btrfs_item_ptr(buf, i,
3229                                             struct btrfs_file_extent_item);
3230                         if (btrfs_file_extent_type(buf, fi) ==
3231                             BTRFS_FILE_EXTENT_INLINE)
3232                                 continue;
3233                         if (btrfs_file_extent_disk_bytenr(buf, fi) == 0)
3234                                 continue;
3235
3236                         data_bytes_allocated +=
3237                                 btrfs_file_extent_disk_num_bytes(buf, fi);
3238                         if (data_bytes_allocated < root->sectorsize) {
3239                                 abort();
3240                         }
3241                         data_bytes_referenced +=
3242                                 btrfs_file_extent_num_bytes(buf, fi);
3243                         ret = add_extent_rec(extent_cache, NULL,
3244                                    btrfs_file_extent_disk_bytenr(buf, fi),
3245                                    btrfs_file_extent_disk_num_bytes(buf, fi),
3246                                    0, 0, 1, 1, 0, 0,
3247                                    btrfs_file_extent_disk_num_bytes(buf, fi));
3248                         add_data_backref(extent_cache,
3249                                 btrfs_file_extent_disk_bytenr(buf, fi),
3250                                 parent, owner, key.objectid, key.offset -
3251                                 btrfs_file_extent_offset(buf, fi), 1, 1,
3252                                 btrfs_file_extent_disk_num_bytes(buf, fi));
3253                         BUG_ON(ret);
3254                 }
3255         } else {
3256                 int level;
3257                 struct btrfs_key first_key;
3258
3259                 first_key.objectid = 0;
3260
3261                 if (nritems > 0)
3262                         btrfs_item_key_to_cpu(buf, &first_key, 0);
3263                 level = btrfs_header_level(buf);
3264                 for (i = 0; i < nritems; i++) {
3265                         u64 ptr = btrfs_node_blockptr(buf, i);
3266                         u32 size = btrfs_level_size(root, level - 1);
3267                         btrfs_node_key_to_cpu(buf, &key, i);
3268                         ret = add_extent_rec(extent_cache, &key,
3269                                              ptr, size, 0, 0, 1, 0, 1, 0,
3270                                              size);
3271                         BUG_ON(ret);
3272
3273                         add_tree_backref(extent_cache, ptr, parent, owner, 1);
3274
3275                         if (level > 1) {
3276                                 add_pending(nodes, seen, ptr, size);
3277                         } else {
3278                                 add_pending(pending, seen, ptr, size);
3279                         }
3280                 }
3281                 btree_space_waste += (BTRFS_NODEPTRS_PER_BLOCK(root) -
3282                                       nritems) * sizeof(struct btrfs_key_ptr);
3283         }
3284         total_btree_bytes += buf->len;
3285         if (fs_root_objectid(btrfs_header_owner(buf)))
3286                 total_fs_tree_bytes += buf->len;
3287         if (btrfs_header_owner(buf) == BTRFS_EXTENT_TREE_OBJECTID)
3288                 total_extent_tree_bytes += buf->len;
3289         if (!found_old_backref &&
3290             btrfs_header_owner(buf) == BTRFS_TREE_RELOC_OBJECTID &&
3291             btrfs_header_backref_rev(buf) == BTRFS_MIXED_BACKREF_REV &&
3292             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC))
3293                 found_old_backref = 1;
3294 out:
3295         free_extent_buffer(buf);
3296         return 0;
3297 }
3298
3299 static int add_root_to_pending(struct extent_buffer *buf,
3300                                struct cache_tree *extent_cache,
3301                                struct cache_tree *pending,
3302                                struct cache_tree *seen,
3303                                struct cache_tree *nodes,
3304                                struct btrfs_key *root_key)
3305 {
3306         if (btrfs_header_level(buf) > 0)
3307                 add_pending(nodes, seen, buf->start, buf->len);
3308         else
3309                 add_pending(pending, seen, buf->start, buf->len);
3310         add_extent_rec(extent_cache, NULL, buf->start, buf->len,
3311                        0, 1, 1, 0, 1, 0, buf->len);
3312
3313         if (root_key->objectid == BTRFS_TREE_RELOC_OBJECTID ||
3314             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
3315                 add_tree_backref(extent_cache, buf->start, buf->start,
3316                                  0, 1);
3317         else
3318                 add_tree_backref(extent_cache, buf->start, 0,
3319                                  root_key->objectid, 1);
3320         return 0;
3321 }
3322
3323 /* as we fix the tree, we might be deleting blocks that
3324  * we're tracking for repair.  This hook makes sure we
3325  * remove any backrefs for blocks as we are fixing them.
3326  */
3327 static int free_extent_hook(struct btrfs_trans_handle *trans,
3328                             struct btrfs_root *root,
3329                             u64 bytenr, u64 num_bytes, u64 parent,
3330                             u64 root_objectid, u64 owner, u64 offset,
3331                             int refs_to_drop)
3332 {
3333         struct extent_record *rec;
3334         struct cache_extent *cache;
3335         int is_data;
3336         struct cache_tree *extent_cache = root->fs_info->fsck_extent_cache;
3337
3338         is_data = owner >= BTRFS_FIRST_FREE_OBJECTID;
3339         cache = find_cache_extent(extent_cache, bytenr, num_bytes);
3340         if (!cache)
3341                 return 0;
3342
3343         rec = container_of(cache, struct extent_record, cache);
3344         if (is_data) {
3345                 struct data_backref *back;
3346                 back = find_data_backref(rec, parent, root_objectid, owner,
3347                                          offset, 1, num_bytes);
3348                 if (!back)
3349                         goto out;
3350                 if (back->node.found_ref) {
3351                         back->found_ref -= refs_to_drop;
3352                         if (rec->refs)
3353                                 rec->refs -= refs_to_drop;
3354                 }
3355                 if (back->node.found_extent_tree) {
3356                         back->num_refs -= refs_to_drop;
3357                         if (rec->extent_item_refs)
3358                                 rec->extent_item_refs -= refs_to_drop;
3359                 }
3360                 if (back->found_ref == 0)
3361                         back->node.found_ref = 0;
3362                 if (back->num_refs == 0)
3363                         back->node.found_extent_tree = 0;
3364
3365                 if (!back->node.found_extent_tree && back->node.found_ref) {
3366                         list_del(&back->node.list);
3367                         free(back);
3368                 }
3369         } else {
3370                 struct tree_backref *back;
3371                 back = find_tree_backref(rec, parent, root_objectid);
3372                 if (!back)
3373                         goto out;
3374                 if (back->node.found_ref) {
3375                         if (rec->refs)
3376                                 rec->refs--;
3377                         back->node.found_ref = 0;
3378                 }
3379                 if (back->node.found_extent_tree) {
3380                         if (rec->extent_item_refs)
3381                                 rec->extent_item_refs--;
3382                         back->node.found_extent_tree = 0;
3383                 }
3384                 if (!back->node.found_extent_tree && back->node.found_ref) {
3385                         list_del(&back->node.list);
3386                         free(back);
3387                 }
3388         }
3389         maybe_free_extent_rec(extent_cache, rec);
3390 out:
3391         return 0;
3392 }
3393
3394 static int delete_extent_records(struct btrfs_trans_handle *trans,
3395                                  struct btrfs_root *root,
3396                                  struct btrfs_path *path,
3397                                  u64 bytenr, u64 new_len)
3398 {
3399         struct btrfs_key key;
3400         struct btrfs_key found_key;
3401         struct extent_buffer *leaf;
3402         int ret;
3403         int slot;
3404
3405
3406         key.objectid = bytenr;
3407         key.type = (u8)-1;
3408         key.offset = (u64)-1;
3409
3410         while(1) {
3411                 ret = btrfs_search_slot(trans, root->fs_info->extent_root,
3412                                         &key, path, 0, 1);
3413                 if (ret < 0)
3414                         break;
3415
3416                 if (ret > 0) {
3417                         ret = 0;
3418                         if (path->slots[0] == 0)
3419                                 break;
3420                         path->slots[0]--;
3421                 }
3422                 ret = 0;
3423
3424                 leaf = path->nodes[0];
3425                 slot = path->slots[0];
3426
3427                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3428                 if (found_key.objectid != bytenr)
3429                         break;
3430
3431                 if (found_key.type != BTRFS_EXTENT_ITEM_KEY &&
3432                     found_key.type != BTRFS_METADATA_ITEM_KEY &&
3433                     found_key.type != BTRFS_TREE_BLOCK_REF_KEY &&
3434                     found_key.type != BTRFS_EXTENT_DATA_REF_KEY &&
3435                     found_key.type != BTRFS_EXTENT_REF_V0_KEY &&
3436                     found_key.type != BTRFS_SHARED_BLOCK_REF_KEY &&
3437                     found_key.type != BTRFS_SHARED_DATA_REF_KEY) {
3438                         btrfs_release_path(NULL, path);
3439                         if (found_key.type == 0) {
3440                                 if (found_key.offset == 0)
3441                                         break;
3442                                 key.offset = found_key.offset - 1;
3443                                 key.type = found_key.type;
3444                         }
3445                         key.type = found_key.type - 1;
3446                         key.offset = (u64)-1;
3447                         continue;
3448                 }
3449
3450                 fprintf(stderr, "repair deleting extent record: key %Lu %u %Lu\n",
3451                         found_key.objectid, found_key.type, found_key.offset);
3452
3453                 ret = btrfs_del_item(trans, root->fs_info->extent_root, path);
3454                 if (ret)
3455                         break;
3456                 btrfs_release_path(NULL, path);
3457
3458                 if (found_key.type == BTRFS_EXTENT_ITEM_KEY ||
3459                     found_key.type == BTRFS_METADATA_ITEM_KEY) {
3460                         u64 bytes = (found_key.type == BTRFS_EXTENT_ITEM_KEY) ?
3461                                 found_key.offset : root->leafsize;
3462
3463                         ret = btrfs_update_block_group(trans, root, bytenr,
3464                                                        bytes, 0, 0);
3465                         if (ret)
3466                                 break;
3467                 }
3468         }
3469
3470         btrfs_release_path(NULL, path);
3471         return ret;
3472 }
3473
3474 /*
3475  * for a single backref, this will allocate a new extent
3476  * and add the backref to it.
3477  */
3478 static int record_extent(struct btrfs_trans_handle *trans,
3479                          struct btrfs_fs_info *info,
3480                          struct btrfs_path *path,
3481                          struct extent_record *rec,
3482                          struct extent_backref *back,
3483                          int allocated, u64 flags)
3484 {
3485         int ret;
3486         struct btrfs_root *extent_root = info->extent_root;
3487         struct extent_buffer *leaf;
3488         struct btrfs_key ins_key;
3489         struct btrfs_extent_item *ei;
3490         struct tree_backref *tback;
3491         struct data_backref *dback;
3492         struct btrfs_tree_block_info *bi;
3493
3494         if (!back->is_data)
3495                 rec->max_size = max_t(u64, rec->max_size,
3496                                     info->extent_root->leafsize);
3497
3498         if (!allocated) {
3499                 u32 item_size = sizeof(*ei);
3500
3501                 if (!back->is_data)
3502                         item_size += sizeof(*bi);
3503
3504                 ins_key.objectid = rec->start;
3505                 ins_key.offset = rec->max_size;
3506                 ins_key.type = BTRFS_EXTENT_ITEM_KEY;
3507
3508                 ret = btrfs_insert_empty_item(trans, extent_root, path,
3509                                         &ins_key, item_size);
3510                 if (ret)
3511                         goto fail;
3512
3513                 leaf = path->nodes[0];
3514                 ei = btrfs_item_ptr(leaf, path->slots[0],
3515                                     struct btrfs_extent_item);
3516
3517                 btrfs_set_extent_refs(leaf, ei, 0);
3518                 btrfs_set_extent_generation(leaf, ei, rec->generation);
3519
3520                 if (back->is_data) {
3521                         btrfs_set_extent_flags(leaf, ei,
3522                                                BTRFS_EXTENT_FLAG_DATA);
3523                 } else {
3524                         struct btrfs_disk_key copy_key;;
3525
3526                         tback = (struct tree_backref *)back;
3527                         bi = (struct btrfs_tree_block_info *)(ei + 1);
3528                         memset_extent_buffer(leaf, 0, (unsigned long)bi,
3529                                              sizeof(*bi));
3530                         memset(&copy_key, 0, sizeof(copy_key));
3531
3532                         copy_key.objectid = le64_to_cpu(rec->info_objectid);
3533                         btrfs_set_tree_block_level(leaf, bi, rec->info_level);
3534                         btrfs_set_tree_block_key(leaf, bi, &copy_key);
3535
3536                         btrfs_set_extent_flags(leaf, ei,
3537                                                BTRFS_EXTENT_FLAG_TREE_BLOCK | flags);
3538                 }
3539
3540                 btrfs_mark_buffer_dirty(leaf);
3541                 ret = btrfs_update_block_group(trans, extent_root, rec->start,
3542                                                rec->max_size, 1, 0);
3543                 if (ret)
3544                         goto fail;
3545                 btrfs_release_path(NULL, path);
3546         }
3547
3548         if (back->is_data) {
3549                 u64 parent;
3550                 int i;
3551
3552                 dback = (struct data_backref *)back;
3553                 if (back->full_backref)
3554                         parent = dback->parent;
3555                 else
3556                         parent = 0;
3557
3558                 for (i = 0; i < dback->found_ref; i++) {
3559                         /* if parent != 0, we're doing a full backref
3560                          * passing BTRFS_FIRST_FREE_OBJECTID as the owner
3561                          * just makes the backref allocator create a data
3562                          * backref
3563                          */
3564                         ret = btrfs_inc_extent_ref(trans, info->extent_root,
3565                                                    rec->start, rec->max_size,
3566                                                    parent,
3567                                                    dback->root,
3568                                                    parent ?
3569                                                    BTRFS_FIRST_FREE_OBJECTID :
3570                                                    dback->owner,
3571                                                    dback->offset);
3572                         if (ret)
3573                                 break;
3574                 }
3575                 fprintf(stderr, "adding new data backref"
3576                                 " on %llu %s %llu owner %llu"
3577                                 " offset %llu found %d\n",
3578                                 (unsigned long long)rec->start,
3579                                 back->full_backref ?
3580                                 "parent" : "root",
3581                                 back->full_backref ?
3582                                 (unsigned long long)parent :
3583                                 (unsigned long long)dback->root,
3584                                 (unsigned long long)dback->owner,
3585                                 (unsigned long long)dback->offset,
3586                                 dback->found_ref);
3587         } else {
3588                 u64 parent;
3589
3590                 tback = (struct tree_backref *)back;
3591                 if (back->full_backref)
3592                         parent = tback->parent;
3593                 else
3594                         parent = 0;
3595
3596                 ret = btrfs_inc_extent_ref(trans, info->extent_root,
3597                                            rec->start, rec->max_size,
3598                                            parent, tback->root, 0, 0);
3599                 fprintf(stderr, "adding new tree backref on "
3600                         "start %llu len %llu parent %llu root %llu\n",
3601                         rec->start, rec->max_size, tback->parent, tback->root);
3602         }
3603         if (ret)
3604                 goto fail;
3605 fail:
3606         btrfs_release_path(NULL, path);
3607         return ret;
3608 }
3609
3610 /*
3611  * when an incorrect extent item is found, this will delete
3612  * all of the existing entries for it and recreate them
3613  * based on what the tree scan found.
3614  */
3615 static int fixup_extent_refs(struct btrfs_trans_handle *trans,
3616                              struct btrfs_fs_info *info,
3617                              struct extent_record *rec)
3618 {
3619         int ret;
3620         struct btrfs_path *path;
3621         struct list_head *cur = rec->backrefs.next;
3622         struct cache_extent *cache;
3623         struct extent_backref *back;
3624         int allocated = 0;
3625         u64 flags = 0;
3626
3627         /* remember our flags for recreating the extent */
3628         ret = btrfs_lookup_extent_info(NULL, info->extent_root, rec->start,
3629                                        rec->max_size, rec->metadata, NULL,
3630                                        &flags);
3631         if (ret < 0)
3632                 flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
3633
3634         path = btrfs_alloc_path();
3635
3636         /* step one, delete all the existing records */
3637         ret = delete_extent_records(trans, info->extent_root, path,
3638                                     rec->start, rec->max_size);
3639
3640         if (ret < 0)
3641                 goto out;
3642
3643         /* was this block corrupt?  If so, don't add references to it */
3644         cache = find_cache_extent(info->corrupt_blocks, rec->start, rec->max_size);
3645         if (cache) {
3646                 ret = 0;
3647                 goto out;
3648         }
3649
3650         /* step two, recreate all the refs we did find */
3651         while(cur != &rec->backrefs) {
3652                 back = list_entry(cur, struct extent_backref, list);
3653                 cur = cur->next;
3654
3655                 /*
3656                  * if we didn't find any references, don't create a
3657                  * new extent record
3658                  */
3659                 if (!back->found_ref)
3660                         continue;
3661
3662                 ret = record_extent(trans, info, path, rec, back, allocated, flags);
3663                 allocated = 1;
3664
3665                 if (ret)
3666                         goto out;
3667         }
3668 out:
3669         btrfs_free_path(path);
3670         return ret;
3671 }
3672
3673 /* right now we only prune from the extent allocation tree */
3674 static int prune_one_block(struct btrfs_trans_handle *trans,
3675                            struct btrfs_fs_info *info,
3676                            struct btrfs_corrupt_block *corrupt)
3677 {
3678         int ret;
3679         struct btrfs_path path;
3680         struct extent_buffer *eb;
3681         u64 found;
3682         int slot;
3683         int nritems;
3684         int level = corrupt->level + 1;
3685
3686         btrfs_init_path(&path);
3687 again:
3688         /* we want to stop at the parent to our busted block */
3689         path.lowest_level = level;
3690
3691         ret = btrfs_search_slot(trans, info->extent_root,
3692                                 &corrupt->key, &path, -1, 1);
3693
3694         if (ret < 0)
3695                 goto out;
3696
3697         eb = path.nodes[level];
3698         if (!eb) {
3699                 ret = -ENOENT;
3700                 goto out;
3701         }
3702
3703         /*
3704          * hopefully the search gave us the block we want to prune,
3705          * lets try that first
3706          */
3707         slot = path.slots[level];
3708         found =  btrfs_node_blockptr(eb, slot);
3709         if (found == corrupt->cache.start)
3710                 goto del_ptr;
3711
3712         nritems = btrfs_header_nritems(eb);
3713
3714         /* the search failed, lets scan this node and hope we find it */
3715         for (slot = 0; slot < nritems; slot++) {
3716                 found =  btrfs_node_blockptr(eb, slot);
3717                 if (found == corrupt->cache.start)
3718                         goto del_ptr;
3719         }
3720         /*
3721          * we couldn't find the bad block.  TODO, search all the nodes for pointers
3722          * to this block
3723          */
3724         if (eb == info->extent_root->node) {
3725                 ret = -ENOENT;
3726                 goto out;
3727         } else {
3728                 level++;
3729                 btrfs_release_path(NULL, &path);
3730                 goto again;
3731         }
3732
3733 del_ptr:
3734         printk("deleting pointer to block %Lu\n", corrupt->cache.start);
3735         ret = btrfs_del_ptr(trans, info->extent_root, &path, level, slot);
3736
3737 out:
3738         btrfs_release_path(NULL, &path);
3739         return ret;
3740 }
3741
3742 static int prune_corrupt_blocks(struct btrfs_trans_handle *trans,
3743                                 struct btrfs_fs_info *info)
3744 {
3745         struct cache_extent *cache;
3746         struct btrfs_corrupt_block *corrupt;
3747
3748         cache = find_first_cache_extent(info->corrupt_blocks, 0);
3749         while (1) {
3750                 if (!cache)
3751                         break;
3752                 corrupt = container_of(cache, struct btrfs_corrupt_block, cache);
3753                 prune_one_block(trans, info, corrupt);
3754                 cache = next_cache_extent(cache);
3755         }
3756         return 0;
3757 }
3758
3759 static void free_corrupt_blocks(struct btrfs_fs_info *info)
3760 {
3761         struct cache_extent *cache;
3762         struct btrfs_corrupt_block *corrupt;
3763
3764         while (1) {
3765                 cache = find_first_cache_extent(info->corrupt_blocks, 0);
3766                 if (!cache)
3767                         break;
3768                 corrupt = container_of(cache, struct btrfs_corrupt_block, cache);
3769                 remove_cache_extent(info->corrupt_blocks, cache);
3770                 free(corrupt);
3771         }
3772 }
3773
3774 static int check_block_group(struct btrfs_trans_handle *trans,
3775                               struct btrfs_fs_info *info,
3776                               struct map_lookup *map,
3777                               int *reinit)
3778 {
3779         struct btrfs_key key;
3780         struct btrfs_path path;
3781         int ret;
3782
3783         key.objectid = map->ce.start;
3784         key.offset = map->ce.size;
3785         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3786
3787         btrfs_init_path(&path);
3788         ret = btrfs_search_slot(NULL, info->extent_root,
3789                                 &key, &path, 0, 0);
3790         btrfs_release_path(NULL, &path);
3791         if (ret <= 0)
3792                 goto out;
3793
3794         ret = btrfs_make_block_group(trans, info->extent_root, 0, map->type,
3795                                BTRFS_FIRST_CHUNK_TREE_OBJECTID,
3796                                key.objectid, key.offset);
3797         *reinit = 1;
3798 out:
3799         return ret;
3800 }
3801
3802 static int check_block_groups(struct btrfs_trans_handle *trans,
3803                               struct btrfs_fs_info *info, int *reinit)
3804 {
3805         struct cache_extent *ce;
3806         struct map_lookup *map;
3807         struct btrfs_mapping_tree *map_tree = &info->mapping_tree;
3808
3809         /* this isn't quite working */
3810         return 0;
3811
3812         ce = find_first_cache_extent(&map_tree->cache_tree, 0);
3813         while (1) {
3814                 if (!ce)
3815                         break;
3816                 map = container_of(ce, struct map_lookup, ce);
3817                 check_block_group(trans, info, map, reinit);
3818                 ce = next_cache_extent(ce);
3819         }
3820         return 0;
3821 }
3822
3823 static int check_extent_refs(struct btrfs_trans_handle *trans,
3824                              struct btrfs_root *root,
3825                              struct cache_tree *extent_cache, int repair)
3826 {
3827         struct extent_record *rec;
3828         struct cache_extent *cache;
3829         int err = 0;
3830         int ret = 0;
3831         int fixed = 0;
3832         int reinit = 0;
3833
3834         if (repair) {
3835                 /*
3836                  * if we're doing a repair, we have to make sure
3837                  * we don't allocate from the problem extents.
3838                  * In the worst case, this will be all the
3839                  * extents in the FS
3840                  */
3841                 cache = find_first_cache_extent(extent_cache, 0);
3842                 while(cache) {
3843                         rec = container_of(cache, struct extent_record, cache);
3844                         btrfs_pin_extent(root->fs_info,
3845                                          rec->start, rec->max_size);
3846                         cache = next_cache_extent(cache);
3847                 }
3848
3849                 /* pin down all the corrupted blocks too */
3850                 cache = find_first_cache_extent(root->fs_info->corrupt_blocks, 0);
3851                 while(cache) {
3852                         rec = container_of(cache, struct extent_record, cache);
3853                         btrfs_pin_extent(root->fs_info,
3854                                          rec->start, rec->max_size);
3855                         cache = next_cache_extent(cache);
3856                 }
3857                 prune_corrupt_blocks(trans, root->fs_info);
3858                 check_block_groups(trans, root->fs_info, &reinit);
3859                 if (reinit)
3860                         btrfs_read_block_groups(root->fs_info->extent_root);
3861         }
3862         while(1) {
3863                 fixed = 0;
3864                 cache = find_first_cache_extent(extent_cache, 0);
3865                 if (!cache)
3866                         break;
3867                 rec = container_of(cache, struct extent_record, cache);
3868                 if (rec->refs != rec->extent_item_refs) {
3869                         fprintf(stderr, "ref mismatch on [%llu %llu] ",
3870                                 (unsigned long long)rec->start,
3871                                 (unsigned long long)rec->nr);
3872                         fprintf(stderr, "extent item %llu, found %llu\n",
3873                                 (unsigned long long)rec->extent_item_refs,
3874                                 (unsigned long long)rec->refs);
3875                         if (!fixed && repair) {
3876                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3877                                 if (ret)
3878                                         goto repair_abort;
3879                                 fixed = 1;
3880                         }
3881                         err = 1;
3882
3883                 }
3884                 if (all_backpointers_checked(rec, 1)) {
3885                         fprintf(stderr, "backpointer mismatch on [%llu %llu]\n",
3886                                 (unsigned long long)rec->start,
3887                                 (unsigned long long)rec->nr);
3888
3889                         if (!fixed && repair) {
3890                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3891                                 if (ret)
3892                                         goto repair_abort;
3893                                 fixed = 1;
3894                         }
3895
3896                         err = 1;
3897                 }
3898                 if (!rec->owner_ref_checked) {
3899                         fprintf(stderr, "owner ref check failed [%llu %llu]\n",
3900                                 (unsigned long long)rec->start,
3901                                 (unsigned long long)rec->nr);
3902                         if (!fixed && repair) {
3903                                 ret = fixup_extent_refs(trans, root->fs_info, rec);
3904                                 if (ret)
3905                                         goto repair_abort;
3906                                 fixed = 1;
3907                         }
3908                         err = 1;
3909                 }
3910
3911                 remove_cache_extent(extent_cache, cache);
3912                 free_all_extent_backrefs(rec);
3913                 free(rec);
3914         }
3915 repair_abort:
3916         if (repair) {
3917                 if (ret) {
3918                         fprintf(stderr, "failed to repair damaged filesystem, aborting\n");
3919                         exit(1);
3920                 } else {
3921                         btrfs_fix_block_accounting(trans, root);
3922                 }
3923                 if (err)
3924                         fprintf(stderr, "repaired damaged extent references\n");
3925                 return ret;
3926         }
3927         return err;
3928 }
3929
3930 static int check_extents(struct btrfs_trans_handle *trans,
3931                          struct btrfs_root *root, int repair)
3932 {
3933         struct cache_tree extent_cache;
3934         struct cache_tree seen;
3935         struct cache_tree pending;
3936         struct cache_tree reada;
3937         struct cache_tree nodes;
3938         struct cache_tree corrupt_blocks;
3939         struct btrfs_path path;
3940         struct btrfs_key key;
3941         struct btrfs_key found_key;
3942         int ret;
3943         u64 last = 0;
3944         struct block_info *bits;
3945         int bits_nr;
3946         struct extent_buffer *leaf;
3947         int slot;
3948         struct btrfs_root_item ri;
3949
3950         cache_tree_init(&extent_cache);
3951         cache_tree_init(&seen);
3952         cache_tree_init(&pending);
3953         cache_tree_init(&nodes);
3954         cache_tree_init(&reada);
3955         cache_tree_init(&corrupt_blocks);
3956
3957         if (repair) {
3958                 root->fs_info->fsck_extent_cache = &extent_cache;
3959                 root->fs_info->free_extent_hook = free_extent_hook;
3960                 root->fs_info->corrupt_blocks = &corrupt_blocks;
3961         }
3962
3963         bits_nr = 1024;
3964         bits = malloc(bits_nr * sizeof(struct block_info));
3965         if (!bits) {
3966                 perror("malloc");
3967                 exit(1);
3968         }
3969
3970         add_root_to_pending(root->fs_info->tree_root->node,
3971                             &extent_cache, &pending, &seen, &nodes,
3972                             &root->fs_info->tree_root->root_key);
3973
3974         add_root_to_pending(root->fs_info->chunk_root->node,
3975                             &extent_cache, &pending, &seen, &nodes,
3976                             &root->fs_info->chunk_root->root_key);
3977
3978         btrfs_init_path(&path);
3979         key.offset = 0;
3980         key.objectid = 0;
3981         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
3982         ret = btrfs_search_slot(NULL, root->fs_info->tree_root,
3983                                         &key, &path, 0, 0);
3984         BUG_ON(ret < 0);
3985         while(1) {
3986                 leaf = path.nodes[0];
3987                 slot = path.slots[0];
3988                 if (slot >= btrfs_header_nritems(path.nodes[0])) {
3989                         ret = btrfs_next_leaf(root, &path);
3990                         if (ret != 0)
3991                                 break;
3992                         leaf = path.nodes[0];
3993                         slot = path.slots[0];
3994                 }
3995                 btrfs_item_key_to_cpu(leaf, &found_key, path.slots[0]);
3996                 if (btrfs_key_type(&found_key) == BTRFS_ROOT_ITEM_KEY) {
3997                         unsigned long offset;
3998                         struct extent_buffer *buf;
3999
4000                         offset = btrfs_item_ptr_offset(leaf, path.slots[0]);
4001                         read_extent_buffer(leaf, &ri, offset, sizeof(ri));
4002                         buf = read_tree_block(root->fs_info->tree_root,
4003                                               btrfs_root_bytenr(&ri),
4004                                               btrfs_level_size(root,
4005                                                btrfs_root_level(&ri)), 0);
4006                         add_root_to_pending(buf, &extent_cache, &pending,
4007                                             &seen, &nodes, &found_key);
4008                         free_extent_buffer(buf);
4009                 }
4010                 path.slots[0]++;
4011         }
4012         btrfs_release_path(root, &path);
4013         while(1) {
4014                 ret = run_next_block(root, bits, bits_nr, &last, &pending,
4015                                      &seen, &reada, &nodes, &extent_cache);
4016                 if (ret != 0)
4017                         break;
4018         }
4019         ret = check_extent_refs(trans, root, &extent_cache, repair);
4020
4021         if (repair) {
4022                 free_corrupt_blocks(root->fs_info);
4023                 root->fs_info->fsck_extent_cache = NULL;
4024                 root->fs_info->free_extent_hook = NULL;
4025                 root->fs_info->corrupt_blocks = NULL;
4026         }
4027
4028         free(bits);
4029         return ret;
4030 }
4031
4032 static struct option long_options[] = {
4033         { "super", 1, NULL, 's' },
4034         { "repair", 0, NULL, 0 },
4035         { "init-csum-tree", 0, NULL, 0 },
4036         { "init-extent-tree", 0, NULL, 0 },
4037         { 0, 0, 0, 0}
4038 };
4039
4040 const char * const cmd_check_usage[] = {
4041         "btrfs check [options] <device>",
4042         "Check an unmounted btrfs filesystem.",
4043         "",
4044         "-s|--super <superblock>     use this superblock copy",
4045         "--repair                    try to repair the filesystem",
4046         "--init-csum-tree            create a new CRC tree",
4047         "--init-extent-tree          create a new extent tree",
4048         NULL
4049 };
4050
4051 int cmd_check(int argc, char **argv)
4052 {
4053         struct cache_tree root_cache;
4054         struct btrfs_root *root;
4055         struct btrfs_fs_info *info;
4056         struct btrfs_trans_handle *trans = NULL;
4057         u64 bytenr = 0;
4058         char uuidbuf[37];
4059         int ret;
4060         int num;
4061         int repair = 0;
4062         int option_index = 0;
4063         int init_csum_tree = 0;
4064         int rw = 0;
4065
4066         while(1) {
4067                 int c;
4068                 c = getopt_long(argc, argv, "as:", long_options,
4069                                 &option_index);
4070                 if (c < 0)
4071                         break;
4072                 switch(c) {
4073                         case 'a': /* ignored */ break;
4074                         case 's':
4075                                 num = atol(optarg);
4076                                 bytenr = btrfs_sb_offset(num);
4077                                 printf("using SB copy %d, bytenr %llu\n", num,
4078                                        (unsigned long long)bytenr);
4079                                 break;
4080                         case '?':
4081                         case 'h':
4082                                 usage(cmd_check_usage);
4083                 }
4084                 if (option_index == 1) {
4085                         printf("enabling repair mode\n");
4086                         repair = 1;
4087                         rw = 1;
4088                 } else if (option_index == 2) {
4089                         printf("Creating a new CRC tree\n");
4090                         init_csum_tree = 1;
4091                         rw = 1;
4092                 }
4093
4094         }
4095         argc = argc - optind;
4096
4097         if (argc != 1)
4098                 usage(cmd_check_usage);
4099
4100         radix_tree_init();
4101         cache_tree_init(&root_cache);
4102
4103         if((ret = check_mounted(argv[optind])) < 0) {
4104                 fprintf(stderr, "Could not check mount status: %s\n", strerror(-ret));
4105                 return ret;
4106         } else if(ret) {
4107                 fprintf(stderr, "%s is currently mounted. Aborting.\n", argv[optind]);
4108                 return -EBUSY;
4109         }
4110
4111         info = open_ctree_fs_info(argv[optind], bytenr, 0, rw, 1);
4112         if (!info) {
4113                 fprintf(stderr, "Couldn't open file system\n");
4114                 return -EIO;
4115         }
4116
4117         uuid_unparse(info->super_copy->fsid, uuidbuf);
4118         printf("Checking filesystem on %s\nUUID: %s\n", argv[optind], uuidbuf);
4119
4120         if (!extent_buffer_uptodate(info->tree_root->node) ||
4121             !extent_buffer_uptodate(info->dev_root->node) ||
4122             !extent_buffer_uptodate(info->extent_root->node) ||
4123             !extent_buffer_uptodate(info->chunk_root->node)) {
4124                 fprintf(stderr, "Critical roots corrupted, unable to fsck the FS\n");
4125                 return -EIO;
4126         }
4127
4128         root = info->fs_root;
4129
4130         fprintf(stderr, "checking extents\n");
4131         if (rw)
4132                 trans = btrfs_start_transaction(root, 1);
4133
4134         if (init_csum_tree) {
4135                 fprintf(stderr, "Reinit crc root\n");
4136                 ret = btrfs_fsck_reinit_root(trans, info->csum_root);
4137                 if (ret) {
4138                         fprintf(stderr, "crc root initialization failed\n");
4139                         return -EIO;
4140                 }
4141                 goto out;
4142         }
4143         ret = check_extents(trans, root, repair);
4144         if (ret)
4145                 fprintf(stderr, "Errors found in extent allocation tree\n");
4146
4147         fprintf(stderr, "checking free space cache\n");
4148         ret = check_space_cache(root);
4149         if (ret)
4150                 goto out;
4151
4152         fprintf(stderr, "checking fs roots\n");
4153         ret = check_fs_roots(root, &root_cache);
4154         if (ret)
4155                 goto out;
4156
4157         fprintf(stderr, "checking csums\n");
4158         ret = check_csums(root);
4159         if (ret)
4160                 goto out;
4161
4162         fprintf(stderr, "checking root refs\n");
4163         ret = check_root_refs(root, &root_cache);
4164 out:
4165         free_root_recs(&root_cache);
4166         if (rw) {
4167                 ret = btrfs_commit_transaction(trans, root);
4168                 if (ret)
4169                         exit(1);
4170         }
4171         close_ctree(root);
4172
4173         if (found_old_backref) { /*
4174                  * there was a disk format change when mixed
4175                  * backref was in testing tree. The old format
4176                  * existed about one week.
4177                  */
4178                 printf("\n * Found old mixed backref format. "
4179                        "The old format is not supported! *"
4180                        "\n * Please mount the FS in readonly mode, "
4181                        "backup data and re-format the FS. *\n\n");
4182                 ret = 1;
4183         }
4184         printf("found %llu bytes used err is %d\n",
4185                (unsigned long long)bytes_used, ret);
4186         printf("total csum bytes: %llu\n",(unsigned long long)total_csum_bytes);
4187         printf("total tree bytes: %llu\n",
4188                (unsigned long long)total_btree_bytes);
4189         printf("total fs tree bytes: %llu\n",
4190                (unsigned long long)total_fs_tree_bytes);
4191         printf("total extent tree bytes: %llu\n",
4192                (unsigned long long)total_extent_tree_bytes);
4193         printf("btree space waste bytes: %llu\n",
4194                (unsigned long long)btree_space_waste);
4195         printf("file data blocks allocated: %llu\n referenced %llu\n",
4196                 (unsigned long long)data_bytes_allocated,
4197                 (unsigned long long)data_bytes_referenced);
4198         printf("%s\n", BTRFS_BUILD_VERSION);
4199         return ret;
4200 }