btrfs-convert.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 600
  20 #define _GNU_SOURCE 1
  21
  22 #include "kerncompat.h"
  23
  24 #include <sys/ioctl.h>
  25 #include <sys/mount.h>
  26 #include <stdio.h>
  27 #include <stdlib.h>
  28 #include <sys/types.h>
  29 #include <sys/stat.h>
  30 #include <sys/acl.h>
  31 #include <fcntl.h>
  32 #include <unistd.h>
  33 #include <uuid/uuid.h>
  34
  35 #include "ctree.h"
  36 #include "disk-io.h"
  37 #include "volumes.h"
  38 #include "transaction.h"
  39 #include "crc32c.h"
  40 #include "utils.h"
  41 #include <ext2fs/ext2_fs.h>
  42 #include <ext2fs/ext2fs.h>
  43 #include <ext2fs/ext2_ext_attr.h>
  44
  45 #define INO_OFFSET (BTRFS_FIRST_FREE_OBJECTID - EXT2_ROOT_INO)
  46 #define STRIPE_LEN (64 * 1024)
  47 #define EXT2_IMAGE_SUBVOL_OBJECTID BTRFS_FIRST_FREE_OBJECTID
  48
  49 /*
  50  * Open Ext2fs in readonly mode, read block allocation bitmap and
  51  * inode bitmap into memory.
  52  */
  53 static int open_ext2fs(const char *name, ext2_filsys *ret_fs)
  54 {
  55         errcode_t ret;
  56         ext2_filsys ext2_fs;
  57         ext2_ino_t ino;
  58         ret = ext2fs_open(name, 0, 0, 0, unix_io_manager, &ext2_fs);
  59         if (ret) {
  60                 fprintf(stderr, "ext2fs_open: %s\n", error_message(ret));
  61                 goto fail;
  62         }
  63         ret = ext2fs_read_inode_bitmap(ext2_fs);
  64         if (ret) {
  65                 fprintf(stderr, "ext2fs_read_inode_bitmap: %s\n",
  66                         error_message(ret));
  67                 goto fail;
  68         }
  69         ret = ext2fs_read_block_bitmap(ext2_fs);
  70         if (ret) {
  71                 fprintf(stderr, "ext2fs_read_block_bitmap: %s\n",
  72                         error_message(ret));
  73                 goto fail;
  74         }
  75         /*
  76          * search each block group for a free inode. this set up
  77          * uninit block/inode bitmaps appropriately.
  78          */
  79         ino = 1;
  80         while (ino <= ext2_fs->super->s_inodes_count) {
  81                 ext2_ino_t foo;
  82                 ext2fs_new_inode(ext2_fs, ino, 0, NULL, &foo);
  83                 ino += EXT2_INODES_PER_GROUP(ext2_fs->super);
  84         }
  85
  86         *ret_fs = ext2_fs;
  87         return 0;
  88 fail:
  89         return -1;
  90 }
  91
  92 static int close_ext2fs(ext2_filsys fs)
  93 {
  94         ext2fs_close(fs);
  95         return 0;
  96 }
  97
  98 static int ext2_alloc_block(ext2_filsys fs, u64 goal, u64 *block_ret)
  99 {
 100         blk_t block;
 101
 102         if (!ext2fs_new_block(fs, goal, NULL, &block)) {
 103                 ext2fs_fast_mark_block_bitmap(fs->block_map, block);
 104                 *block_ret = block;
 105                 return 0;
 106         }
 107         return -ENOSPC;
 108 }
 109
 110 static int ext2_free_block(ext2_filsys fs, u64 block)
 111 {
 112         BUG_ON(block != (blk_t)block);
 113         ext2fs_fast_unmark_block_bitmap(fs->block_map, block);
 114         return 0;
 115 }
 116
 117 static int cache_free_extents(struct btrfs_root *root, ext2_filsys ext2_fs)
 118
 119 {
 120         int i, ret = 0;
 121         blk_t block;
 122         u64 bytenr;
 123         u64 blocksize = ext2_fs->blocksize;
 124
 125         block = ext2_fs->super->s_first_data_block;
 126         for (; block < ext2_fs->super->s_blocks_count; block++) {
 127                 if (ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
 128                         continue;
 129                 bytenr = block * blocksize;
 130                 ret = set_extent_dirty(&root->fs_info->free_space_cache,
 131                                        bytenr, bytenr + blocksize - 1, 0);
 132                 BUG_ON(ret);
 133         }
 134
 135         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 136                 bytenr = btrfs_sb_offset(i);
 137                 bytenr &= ~((u64)STRIPE_LEN - 1);
 138                 if (bytenr >= blocksize * ext2_fs->super->s_blocks_count)
 139                         break;
 140                 clear_extent_dirty(&root->fs_info->free_space_cache, bytenr,
 141                                    bytenr + STRIPE_LEN - 1, 0);
 142         }
 143
 144         clear_extent_dirty(&root->fs_info->free_space_cache,
 145                            0, BTRFS_SUPER_INFO_OFFSET - 1, 0);
 146
 147         return 0;
 148 }
 149
 150 static int custom_alloc_extent(struct btrfs_root *root, u64 num_bytes,
 151                                u64 hint_byte, struct btrfs_key *ins)
 152 {
 153         u64 start;
 154         u64 end;
 155         u64 last = hint_byte;
 156         int ret;
 157         int wrapped = 0;
 158         struct btrfs_block_group_cache *cache;
 159
 160         while(1) {
 161                 ret = find_first_extent_bit(&root->fs_info->free_space_cache,
 162                                             last, &start, &end, EXTENT_DIRTY);
 163                 if (ret) {
 164                         if (wrapped++ == 0) {
 165                                 last = 0;
 166                                 continue;
 167                         } else {
 168                                 goto fail;
 169                         }
 170                 }
 171
 172                 start = max(last, start);
 173                 last = end + 1;
 174                 if (last - start < num_bytes)
 175                         continue;
 176
 177                 last = start + num_bytes;
 178                 if (test_range_bit(&root->fs_info->pinned_extents,
 179                                    start, last - 1, EXTENT_DIRTY, 0))
 180                         continue;
 181
 182                 cache = btrfs_lookup_block_group(root->fs_info, start);
 183                 BUG_ON(!cache);
 184                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM ||
 185                     last > cache->key.objectid + cache->key.offset) {
 186                         last = cache->key.objectid + cache->key.offset;
 187                         continue;
 188                 }
 189
 190                 clear_extent_dirty(&root->fs_info->free_space_cache,
 191                                    start, start + num_bytes - 1, 0);
 192
 193                 ins->objectid = start;
 194                 ins->offset = num_bytes;
 195                 ins->type = BTRFS_EXTENT_ITEM_KEY;
 196                 return 0;
 197         }
 198 fail:
 199         fprintf(stderr, "not enough free space\n");
 200         return -ENOSPC;
 201 }
 202
 203 static int intersect_with_sb(u64 bytenr, u64 num_bytes)
 204 {
 205         int i;
 206         u64 offset;
 207
 208         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 209                 offset = btrfs_sb_offset(i);
 210                 offset &= ~((u64)STRIPE_LEN - 1);
 211
 212                 if (bytenr < offset + STRIPE_LEN &&
 213                     bytenr + num_bytes > offset)
 214                         return 1;
 215         }
 216         return 0;
 217 }
 218
 219 static int custom_free_extent(struct btrfs_root *root, u64 bytenr,
 220                               u64 num_bytes)
 221 {
 222         return intersect_with_sb(bytenr, num_bytes);
 223 }
 224
 225 static struct btrfs_extent_ops extent_ops = {
 226         .alloc_extent = custom_alloc_extent,
 227         .free_extent = custom_free_extent,
 228 };
 229
 230 struct dir_iterate_data {
 231         struct btrfs_trans_handle *trans;
 232         struct btrfs_root *root;
 233         struct btrfs_inode_item *inode;
 234         u64 objectid;
 235         u64 index_cnt;
 236         u64 parent;
 237         int errcode;
 238 };
 239
 240 static u8 filetype_conversion_table[EXT2_FT_MAX] = {
 241         [EXT2_FT_UNKNOWN]       = BTRFS_FT_UNKNOWN,
 242         [EXT2_FT_REG_FILE]      = BTRFS_FT_REG_FILE,
 243         [EXT2_FT_DIR]           = BTRFS_FT_DIR,
 244         [EXT2_FT_CHRDEV]        = BTRFS_FT_CHRDEV,
 245         [EXT2_FT_BLKDEV]        = BTRFS_FT_BLKDEV,
 246         [EXT2_FT_FIFO]          = BTRFS_FT_FIFO,
 247         [EXT2_FT_SOCK]          = BTRFS_FT_SOCK,
 248         [EXT2_FT_SYMLINK]       = BTRFS_FT_SYMLINK,
 249 };
 250
 251 static int dir_iterate_proc(ext2_ino_t dir, int entry,
 252                             struct ext2_dir_entry *dirent,
 253                             int offset, int blocksize,
 254                             char *buf,void *priv_data)
 255 {
 256         int ret;
 257         int file_type;
 258         u64 objectid;
 259         u64 inode_size;
 260         char dotdot[] = "..";
 261         struct btrfs_key location;
 262         struct dir_iterate_data *idata = (struct dir_iterate_data *)priv_data;
 263         int name_len;
 264
 265         name_len = dirent->name_len & 0xFF;
 266
 267         objectid = dirent->inode + INO_OFFSET;
 268         if (!strncmp(dirent->name, dotdot, name_len)) {
 269                 if (name_len == 2) {
 270                         BUG_ON(idata->parent != 0);
 271                         idata->parent = objectid;
 272                 }
 273                 return 0;
 274         }
 275         if (dirent->inode < EXT2_GOOD_OLD_FIRST_INO)
 276                 return 0;
 277
 278         location.objectid = objectid;
 279         location.offset = 0;
 280         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
 281
 282         file_type = dirent->name_len >> 8;
 283         BUG_ON(file_type > EXT2_FT_SYMLINK);
 284         ret = btrfs_insert_dir_item(idata->trans, idata->root,
 285                                     dirent->name, name_len,
 286                                     idata->objectid, &location,
 287                                     filetype_conversion_table[file_type],
 288                                     idata->index_cnt);
 289         if (ret)
 290                 goto fail;
 291         ret = btrfs_insert_inode_ref(idata->trans, idata->root,
 292                                      dirent->name, name_len,
 293                                      objectid, idata->objectid,
 294                                      idata->index_cnt);
 295         if (ret)
 296                 goto fail;
 297         idata->index_cnt++;
 298         inode_size = btrfs_stack_inode_size(idata->inode) +
 299                      name_len * 2;
 300         btrfs_set_stack_inode_size(idata->inode, inode_size);
 301         return 0;
 302 fail:
 303         idata->errcode = ret;
 304         return BLOCK_ABORT;
 305 }
 306
 307 static int create_dir_entries(struct btrfs_trans_handle *trans,
 308                               struct btrfs_root *root, u64 objectid,
 309                               struct btrfs_inode_item *btrfs_inode,
 310                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 311 {
 312         int ret;
 313         errcode_t err;
 314         struct dir_iterate_data data = {
 315                 .trans          = trans,
 316                 .root           = root,
 317                 .inode          = btrfs_inode,
 318                 .objectid       = objectid,
 319                 .index_cnt      = 2,
 320                 .parent         = 0,
 321                 .errcode        = 0,
 322         };
 323
 324         err = ext2fs_dir_iterate2(ext2_fs, ext2_ino, 0, NULL,
 325                                   dir_iterate_proc, &data);
 326         if (err)
 327                 goto error;
 328         ret = data.errcode;
 329         if (ret == 0 && data.parent == objectid) {
 330                 ret = btrfs_insert_inode_ref(trans, root, "..", 2,
 331                                              objectid, objectid, 0);
 332         }
 333         return ret;
 334 error:
 335         fprintf(stderr, "ext2fs_dir_iterate2: %s\n", error_message(err));
 336         return -1;
 337 }
 338
 339 static int read_disk_extent(struct btrfs_root *root, u64 bytenr,
 340                             u32 num_bytes, char *buffer)
 341 {
 342         int ret;
 343         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
 344
 345         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, bytenr);
 346         if (ret != num_bytes)
 347                 goto fail;
 348         ret = 0;
 349 fail:
 350         if (ret > 0)
 351                 ret = -1;
 352         return ret;
 353 }
 354
 355 static int csum_disk_extent(struct btrfs_trans_handle *trans,
 356                             struct btrfs_root *root,
 357                             u64 disk_bytenr, u64 num_bytes)
 358 {
 359         u32 blocksize = root->sectorsize;
 360         u64 offset;
 361         char *buffer;
 362         int ret = 0;
 363
 364         buffer = malloc(blocksize);
 365         if (!buffer)
 366                 return -ENOMEM;
 367         for (offset = 0; offset < num_bytes; offset += blocksize) {
 368                 ret = read_disk_extent(root, disk_bytenr + offset,
 369                                         blocksize, buffer);
 370                 if (ret)
 371                         break;
 372                 ret = btrfs_csum_file_block(trans,
 373                                             root->fs_info->csum_root,
 374                                             disk_bytenr + num_bytes,
 375                                             disk_bytenr + offset,
 376                                             buffer, blocksize);
 377                 if (ret)
 378                         break;
 379         }
 380         free(buffer);
 381         return ret;
 382 }
 383
 384 static int record_file_blocks(struct btrfs_trans_handle *trans,
 385                               struct btrfs_root *root, u64 objectid,
 386                               struct btrfs_inode_item *inode,
 387                               u64 file_block, u64 disk_block,
 388                               u64 num_blocks, int checksum)
 389 {
 390         int ret;
 391         u64 file_pos = file_block * root->sectorsize;
 392         u64 disk_bytenr = disk_block * root->sectorsize;
 393         u64 num_bytes = num_blocks * root->sectorsize;
 394         ret = btrfs_record_file_extent(trans, root, objectid, inode, file_pos,
 395                                         disk_bytenr, num_bytes);
 396
 397         if (ret || !checksum)
 398                 return ret;
 399
 400         return csum_disk_extent(trans, root, disk_bytenr, num_bytes);
 401 }
 402
 403 struct blk_iterate_data {
 404         struct btrfs_trans_handle *trans;
 405         struct btrfs_root *root;
 406         struct btrfs_inode_item *inode;
 407         u64 objectid;
 408         u64 first_block;
 409         u64 disk_block;
 410         u64 num_blocks;
 411         u64 boundary;
 412         int checksum;
 413         int errcode;
 414 };
 415
 416 static int block_iterate_proc(ext2_filsys ext2_fs,
 417                               u64 disk_block, u64 file_block,
 418                               struct blk_iterate_data *idata)
 419 {
 420         int ret;
 421         int sb_region;
 422         int do_barrier;
 423         struct btrfs_root *root = idata->root;
 424         struct btrfs_trans_handle *trans = idata->trans;
 425         struct btrfs_block_group_cache *cache;
 426         u64 bytenr = disk_block * root->sectorsize;
 427
 428         sb_region = intersect_with_sb(bytenr, root->sectorsize);
 429         do_barrier = sb_region || disk_block >= idata->boundary;
 430         if ((idata->num_blocks > 0 && do_barrier) ||
 431             (file_block > idata->first_block + idata->num_blocks) ||
 432             (disk_block != idata->disk_block + idata->num_blocks)) {
 433                 if (idata->num_blocks > 0) {
 434                         ret = record_file_blocks(trans, root, idata->objectid,
 435                                         idata->inode, idata->first_block,
 436                                         idata->disk_block, idata->num_blocks,
 437                                         idata->checksum);
 438                         if (ret)
 439                                 goto fail;
 440                         idata->first_block += idata->num_blocks;
 441                         idata->num_blocks = 0;
 442                 }
 443                 if (file_block > idata->first_block) {
 444                         ret = record_file_blocks(trans, root, idata->objectid,
 445                                         idata->inode, idata->first_block,
 446                                         0, file_block - idata->first_block,
 447                                         idata->checksum);
 448                         if (ret)
 449                                 goto fail;
 450                 }
 451
 452                 if (sb_region) {
 453                         bytenr += STRIPE_LEN - 1;
 454                         bytenr &= ~((u64)STRIPE_LEN - 1);
 455                 } else {
 456                         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
 457                         BUG_ON(!cache);
 458                         bytenr = cache->key.objectid + cache->key.offset;
 459                 }
 460
 461                 idata->first_block = file_block;
 462                 idata->disk_block = disk_block;
 463                 idata->boundary = bytenr / root->sectorsize;
 464         }
 465         idata->num_blocks++;
 466         return 0;
 467 fail:
 468         idata->errcode = ret;
 469         return BLOCK_ABORT;
 470 }
 471
 472 static int __block_iterate_proc(ext2_filsys fs, blk_t *blocknr,
 473                                 e2_blkcnt_t blockcnt, blk_t ref_block,
 474                                 int ref_offset, void *priv_data)
 475 {
 476         struct blk_iterate_data *idata;
 477         idata = (struct blk_iterate_data *)priv_data;
 478         return block_iterate_proc(fs, *blocknr, blockcnt, idata);
 479 }
 480
 481 /*
 482  * traverse file's data blocks, record these data blocks as file extents.
 483  */
 484 static int create_file_extents(struct btrfs_trans_handle *trans,
 485                                struct btrfs_root *root, u64 objectid,
 486                                struct btrfs_inode_item *btrfs_inode,
 487                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 488                                int datacsum, int packing)
 489 {
 490         int ret;
 491         char *buffer = NULL;
 492         errcode_t err;
 493         u32 last_block;
 494         u32 sectorsize = root->sectorsize;
 495         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 496         struct blk_iterate_data data = {
 497                 .trans          = trans,
 498                 .root           = root,
 499                 .inode          = btrfs_inode,
 500                 .objectid       = objectid,
 501                 .first_block    = 0,
 502                 .disk_block     = 0,
 503                 .num_blocks     = 0,
 504                 .boundary       = (u64)-1,
 505                 .checksum       = datacsum,
 506                 .errcode        = 0,
 507         };
 508         err = ext2fs_block_iterate2(ext2_fs, ext2_ino, BLOCK_FLAG_DATA_ONLY,
 509                                     NULL, __block_iterate_proc, &data);
 510         if (err)
 511                 goto error;
 512         ret = data.errcode;
 513         if (ret)
 514                 goto fail;
 515         if (packing && data.first_block == 0 && data.num_blocks > 0 &&
 516             inode_size <= BTRFS_MAX_INLINE_DATA_SIZE(root)) {
 517                 u64 num_bytes = data.num_blocks * sectorsize;
 518                 u64 disk_bytenr = data.disk_block * sectorsize;
 519                 u64 nbytes;
 520
 521                 buffer = malloc(num_bytes);
 522                 if (!buffer)
 523                         return -ENOMEM;
 524                 ret = read_disk_extent(root, disk_bytenr, num_bytes, buffer);
 525                 if (ret)
 526                         goto fail;
 527                 if (num_bytes > inode_size)
 528                         num_bytes = inode_size;
 529                 ret = btrfs_insert_inline_extent(trans, root, objectid,
 530                                                  0, buffer, num_bytes);
 531                 if (ret)
 532                         goto fail;
 533                 nbytes = btrfs_stack_inode_nbytes(btrfs_inode) + num_bytes;
 534                 btrfs_set_stack_inode_nbytes(btrfs_inode, nbytes);
 535         } else if (data.num_blocks > 0) {
 536                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 537                                          data.first_block, data.disk_block,
 538                                          data.num_blocks, data.checksum);
 539                 if (ret)
 540                         goto fail;
 541         }
 542         data.first_block += data.num_blocks;
 543         last_block = (inode_size + sectorsize - 1) / sectorsize;
 544         if (last_block > data.first_block) {
 545                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 546                                          data.first_block, 0, last_block -
 547                                          data.first_block, data.checksum);
 548         }
 549 fail:
 550         free(buffer);
 551         return ret;
 552 error:
 553         fprintf(stderr, "ext2fs_block_iterate2: %s\n", error_message(err));
 554         return -1;
 555 }
 556
 557 static int create_symbol_link(struct btrfs_trans_handle *trans,
 558                               struct btrfs_root *root, u64 objectid,
 559                               struct btrfs_inode_item *btrfs_inode,
 560                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 561                               struct ext2_inode *ext2_inode)
 562 {
 563         int ret;
 564         char *pathname;
 565         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 566         if (ext2fs_inode_data_blocks(ext2_fs, ext2_inode)) {
 567                 btrfs_set_stack_inode_size(btrfs_inode, inode_size + 1);
 568                 ret = create_file_extents(trans, root, objectid, btrfs_inode,
 569                                           ext2_fs, ext2_ino, 1, 1);
 570                 btrfs_set_stack_inode_size(btrfs_inode, inode_size);
 571                 return ret;
 572         }
 573
 574         pathname = (char *)&(ext2_inode->i_block[0]);
 575         BUG_ON(pathname[inode_size] != 0);
 576         ret = btrfs_insert_inline_extent(trans, root, objectid, 0,
 577                                          pathname, inode_size + 1);
 578         btrfs_set_stack_inode_nbytes(btrfs_inode, inode_size + 1);
 579         return ret;
 580 }
 581
 582 /*
 583  * Following xattr/acl related codes are based on codes in
 584  * fs/ext3/xattr.c and fs/ext3/acl.c
 585  */
 586 #define EXT2_XATTR_BHDR(ptr) ((struct ext2_ext_attr_header *)(ptr))
 587 #define EXT2_XATTR_BFIRST(ptr) \
 588         ((struct ext2_ext_attr_entry *)(EXT2_XATTR_BHDR(ptr) + 1))
 589 #define EXT2_XATTR_IHDR(inode) \
 590         ((struct ext2_ext_attr_header *) ((void *)(inode) + \
 591                 EXT2_GOOD_OLD_INODE_SIZE + (inode)->i_extra_isize))
 592 #define EXT2_XATTR_IFIRST(inode) \
 593         ((struct ext2_ext_attr_entry *) ((void *)EXT2_XATTR_IHDR(inode) + \
 594                 sizeof(EXT2_XATTR_IHDR(inode)->h_magic)))
 595
 596 static int ext2_xattr_check_names(struct ext2_ext_attr_entry *entry,
 597                                   const void *end)
 598 {
 599         struct ext2_ext_attr_entry *next;
 600
 601         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 602                 next = EXT2_EXT_ATTR_NEXT(entry);
 603                 if ((void *)next >= end)
 604                         return -EIO;
 605                 entry = next;
 606         }
 607         return 0;
 608 }
 609
 610 static int ext2_xattr_check_block(const char *buf, size_t size)
 611 {
 612         int error;
 613         struct ext2_ext_attr_header *header = EXT2_XATTR_BHDR(buf);
 614
 615         if (header->h_magic != EXT2_EXT_ATTR_MAGIC ||
 616             header->h_blocks != 1)
 617                 return -EIO;
 618         error = ext2_xattr_check_names(EXT2_XATTR_BFIRST(buf), buf + size);
 619         return error;
 620 }
 621
 622 static int ext2_xattr_check_entry(struct ext2_ext_attr_entry *entry,
 623                                   size_t size)
 624 {
 625         size_t value_size = entry->e_value_size;
 626
 627         if (entry->e_value_block != 0 || value_size > size ||
 628             entry->e_value_offs + value_size > size)
 629                 return -EIO;
 630         return 0;
 631 }
 632
 633 #define EXT2_ACL_VERSION        0x0001
 634
 635 typedef struct {
 636         __le16          e_tag;
 637         __le16          e_perm;
 638         __le32          e_id;
 639 } ext2_acl_entry;
 640
 641 typedef struct {
 642         __le16          e_tag;
 643         __le16          e_perm;
 644 } ext2_acl_entry_short;
 645
 646 typedef struct {
 647         __le32          a_version;
 648 } ext2_acl_header;
 649
 650 static inline int ext2_acl_count(size_t size)
 651 {
 652         ssize_t s;
 653         size -= sizeof(ext2_acl_header);
 654         s = size - 4 * sizeof(ext2_acl_entry_short);
 655         if (s < 0) {
 656                 if (size % sizeof(ext2_acl_entry_short))
 657                         return -1;
 658                 return size / sizeof(ext2_acl_entry_short);
 659         } else {
 660                 if (s % sizeof(ext2_acl_entry))
 661                         return -1;
 662                 return s / sizeof(ext2_acl_entry) + 4;
 663         }
 664 }
 665
 666 #define ACL_EA_VERSION          0x0002
 667
 668 typedef struct {
 669         __le16          e_tag;
 670         __le16          e_perm;
 671         __le32          e_id;
 672 } acl_ea_entry;
 673
 674 typedef struct {
 675         __le32          a_version;
 676         acl_ea_entry    a_entries[0];
 677 } acl_ea_header;
 678
 679 static inline size_t acl_ea_size(int count)
 680 {
 681         return sizeof(acl_ea_header) + count * sizeof(acl_ea_entry);
 682 }
 683
 684 static int ext2_acl_to_xattr(void *dst, const void *src,
 685                              size_t dst_size, size_t src_size)
 686 {
 687         int i, count;
 688         const void *end = src + src_size;
 689         acl_ea_header *ext_acl = (acl_ea_header *)dst;
 690         acl_ea_entry *dst_entry = ext_acl->a_entries;
 691         ext2_acl_entry *src_entry;
 692
 693         if (src_size < sizeof(ext2_acl_header))
 694                 goto fail;
 695         if (((ext2_acl_header *)src)->a_version !=
 696             cpu_to_le32(EXT2_ACL_VERSION))
 697                 goto fail;
 698         src += sizeof(ext2_acl_header);
 699         count = ext2_acl_count(src_size);
 700         if (count <= 0)
 701                 goto fail;
 702
 703         BUG_ON(dst_size < acl_ea_size(count));
 704         ext_acl->a_version = cpu_to_le32(ACL_EA_VERSION);
 705         for (i = 0; i < count; i++, dst_entry++) {
 706                 src_entry = (ext2_acl_entry *)src;
 707                 if (src + sizeof(ext2_acl_entry_short) > end)
 708                         goto fail;
 709                 dst_entry->e_tag = src_entry->e_tag;
 710                 dst_entry->e_perm = src_entry->e_perm;
 711                 switch (le16_to_cpu(src_entry->e_tag)) {
 712                 case ACL_USER_OBJ:
 713                 case ACL_GROUP_OBJ:
 714                 case ACL_MASK:
 715                 case ACL_OTHER:
 716                         src += sizeof(ext2_acl_entry_short);
 717                         dst_entry->e_id = cpu_to_le32(ACL_UNDEFINED_ID);
 718                         break;
 719                 case ACL_USER:
 720                 case ACL_GROUP:
 721                         src += sizeof(ext2_acl_entry);
 722                         if (src > end)
 723                                 goto fail;
 724                         dst_entry->e_id = src_entry->e_id;
 725                         break;
 726                 default:
 727                         goto fail;
 728                 }
 729         }
 730         if (src != end)
 731                 goto fail;
 732         return 0;
 733 fail:
 734         return -EINVAL;
 735 }
 736
 737 static char *xattr_prefix_table[] = {
 738         [1] =   "user.",
 739         [2] =   "system.posix_acl_access",
 740         [3] =   "system.posix_acl_default",
 741         [4] =   "trusted.",
 742         [6] =   "security.",
 743 };
 744
 745 static int copy_single_xattr(struct btrfs_trans_handle *trans,
 746                              struct btrfs_root *root, u64 objectid,
 747                              struct ext2_ext_attr_entry *entry,
 748                              const void *data, u32 datalen)
 749 {
 750         int ret = 0;
 751         int name_len;
 752         int name_index;
 753         void *databuf = NULL;
 754         char namebuf[XATTR_NAME_MAX + 1];
 755
 756         name_index = entry->e_name_index;
 757         if (name_index >= ARRAY_SIZE(xattr_prefix_table) ||
 758             xattr_prefix_table[name_index] == NULL)
 759                 return -EOPNOTSUPP;
 760         name_len = strlen(xattr_prefix_table[name_index]) +
 761                    entry->e_name_len;
 762         if (name_len >= sizeof(namebuf))
 763                 return -ERANGE;
 764
 765         if (name_index == 2 || name_index == 3) {
 766                 size_t bufsize = acl_ea_size(ext2_acl_count(datalen));
 767                 databuf = malloc(bufsize);
 768                 if (!databuf)
 769                        return -ENOMEM;
 770                 ret = ext2_acl_to_xattr(databuf, data, bufsize, datalen);
 771                 if (ret)
 772                         goto out;
 773                 data = databuf;
 774                 datalen = bufsize;
 775         }
 776         strncpy(namebuf, xattr_prefix_table[name_index], XATTR_NAME_MAX);
 777         strncat(namebuf, EXT2_EXT_ATTR_NAME(entry), entry->e_name_len);
 778         if (name_len + datalen > BTRFS_LEAF_DATA_SIZE(root) -
 779             sizeof(struct btrfs_item) - sizeof(struct btrfs_dir_item)) {
 780                 fprintf(stderr, "skip large xattr on inode %Lu name %.*s\n",
 781                         objectid - INO_OFFSET, name_len, namebuf);
 782                 goto out;
 783         }
 784         ret = btrfs_insert_xattr_item(trans, root, namebuf, name_len,
 785                                       data, datalen, objectid);
 786 out:
 787         free(databuf);
 788         return ret;
 789 }
 790
 791 static int copy_extended_attrs(struct btrfs_trans_handle *trans,
 792                                struct btrfs_root *root, u64 objectid,
 793                                struct btrfs_inode_item *btrfs_inode,
 794                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 795 {
 796         int ret = 0;
 797         int inline_ea = 0;
 798         errcode_t err;
 799         u32 datalen;
 800         u32 block_size = ext2_fs->blocksize;
 801         u32 inode_size = EXT2_INODE_SIZE(ext2_fs->super);
 802         struct ext2_inode_large *ext2_inode;
 803         struct ext2_ext_attr_entry *entry;
 804         void *data;
 805         char *buffer = NULL;
 806         char inode_buf[EXT2_GOOD_OLD_INODE_SIZE];
 807
 808         if (inode_size <= EXT2_GOOD_OLD_INODE_SIZE) {
 809                 ext2_inode = (struct ext2_inode_large *)inode_buf;
 810         } else {
 811                 ext2_inode = (struct ext2_inode_large *)malloc(inode_size);
 812                 if (!ext2_inode)
 813                        return -ENOMEM;
 814         }
 815         err = ext2fs_read_inode_full(ext2_fs, ext2_ino, (void *)ext2_inode,
 816                                      inode_size);
 817         if (err) {
 818                 fprintf(stderr, "ext2fs_read_inode_full: %s\n",
 819                         error_message(err));
 820                 ret = -1;
 821                 goto out;
 822         }
 823
 824         if (ext2_ino > ext2_fs->super->s_first_ino &&
 825             inode_size > EXT2_GOOD_OLD_INODE_SIZE) {
 826                 if (EXT2_GOOD_OLD_INODE_SIZE +
 827                     ext2_inode->i_extra_isize > inode_size) {
 828                         ret = -EIO;
 829                         goto out;
 830                 }
 831                 if (ext2_inode->i_extra_isize != 0 &&
 832                     EXT2_XATTR_IHDR(ext2_inode)->h_magic ==
 833                     EXT2_EXT_ATTR_MAGIC) {
 834                         inline_ea = 1;
 835                 }
 836         }
 837         if (inline_ea) {
 838                 int total;
 839                 void *end = (void *)ext2_inode + inode_size;
 840                 entry = EXT2_XATTR_IFIRST(ext2_inode);
 841                 total = end - (void *)entry;
 842                 ret = ext2_xattr_check_names(entry, end);
 843                 if (ret)
 844                         goto out;
 845                 while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 846                         ret = ext2_xattr_check_entry(entry, total);
 847                         if (ret)
 848                                 goto out;
 849                         data = (void *)EXT2_XATTR_IFIRST(ext2_inode) +
 850                                 entry->e_value_offs;
 851                         datalen = entry->e_value_size;
 852                         ret = copy_single_xattr(trans, root, objectid,
 853                                                 entry, data, datalen);
 854                         if (ret)
 855                                 goto out;
 856                         entry = EXT2_EXT_ATTR_NEXT(entry);
 857                 }
 858         }
 859
 860         if (ext2_inode->i_file_acl == 0)
 861                 goto out;
 862
 863         buffer = malloc(block_size);
 864         if (!buffer) {
 865                 ret = -ENOMEM;
 866                 goto out;
 867         }
 868         err = ext2fs_read_ext_attr(ext2_fs, ext2_inode->i_file_acl, buffer);
 869         if (err) {
 870                 fprintf(stderr, "ext2fs_read_ext_attr: %s\n",
 871                         error_message(err));
 872                 ret = -1;
 873                 goto out;
 874         }
 875         ret = ext2_xattr_check_block(buffer, block_size);
 876         if (ret)
 877                 goto out;
 878
 879         entry = EXT2_XATTR_BFIRST(buffer);
 880         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 881                 ret = ext2_xattr_check_entry(entry, block_size);
 882                 if (ret)
 883                         goto out;
 884                 data = buffer + entry->e_value_offs;
 885                 datalen = entry->e_value_size;
 886                 ret = copy_single_xattr(trans, root, objectid,
 887                                         entry, data, datalen);
 888                 if (ret)
 889                         goto out;
 890                 entry = EXT2_EXT_ATTR_NEXT(entry);
 891         }
 892 out:
 893         free(buffer);
 894         if ((void *)ext2_inode != inode_buf)
 895                 free(ext2_inode);
 896         return ret;
 897 }
 898 #define MINORBITS       20
 899 #define MKDEV(ma, mi)   (((ma) << MINORBITS) | (mi))
 900
 901 static inline dev_t old_decode_dev(u16 val)
 902 {
 903         return MKDEV((val >> 8) & 255, val & 255);
 904 }
 905
 906 static inline dev_t new_decode_dev(u32 dev)
 907 {
 908         unsigned major = (dev & 0xfff00) >> 8;
 909         unsigned minor = (dev & 0xff) | ((dev >> 12) & 0xfff00);
 910         return MKDEV(major, minor);
 911 }
 912
 913 static int copy_inode_item(struct btrfs_inode_item *dst,
 914                            struct ext2_inode *src, u32 blocksize)
 915 {
 916         btrfs_set_stack_inode_generation(dst, 1);
 917         btrfs_set_stack_inode_size(dst, src->i_size);
 918         btrfs_set_stack_inode_nbytes(dst, 0);
 919         btrfs_set_stack_inode_block_group(dst, 0);
 920         btrfs_set_stack_inode_nlink(dst, src->i_links_count);
 921         btrfs_set_stack_inode_uid(dst, src->i_uid | (src->i_uid_high << 16));
 922         btrfs_set_stack_inode_gid(dst, src->i_gid | (src->i_gid_high << 16));
 923         btrfs_set_stack_inode_mode(dst, src->i_mode);
 924         btrfs_set_stack_inode_rdev(dst, 0);
 925         btrfs_set_stack_inode_flags(dst, 0);
 926         btrfs_set_stack_timespec_sec(&dst->atime, src->i_atime);
 927         btrfs_set_stack_timespec_nsec(&dst->atime, 0);
 928         btrfs_set_stack_timespec_sec(&dst->ctime, src->i_ctime);
 929         btrfs_set_stack_timespec_nsec(&dst->ctime, 0);
 930         btrfs_set_stack_timespec_sec(&dst->mtime, src->i_mtime);
 931         btrfs_set_stack_timespec_nsec(&dst->mtime, 0);
 932         btrfs_set_stack_timespec_sec(&dst->otime, 0);
 933         btrfs_set_stack_timespec_nsec(&dst->otime, 0);
 934
 935         if (S_ISDIR(src->i_mode)) {
 936                 btrfs_set_stack_inode_size(dst, 0);
 937                 btrfs_set_stack_inode_nlink(dst, 1);
 938         }
 939         if (S_ISREG(src->i_mode)) {
 940                 btrfs_set_stack_inode_size(dst, (u64)src->i_size_high << 32 |
 941                                            (u64)src->i_size);
 942         }
 943         if (!S_ISREG(src->i_mode) && !S_ISDIR(src->i_mode) &&
 944             !S_ISLNK(src->i_mode)) {
 945                 if (src->i_block[0]) {
 946                         btrfs_set_stack_inode_rdev(dst,
 947                                 old_decode_dev(src->i_block[0]));
 948                 } else {
 949                         btrfs_set_stack_inode_rdev(dst,
 950                                 new_decode_dev(src->i_block[1]));
 951                 }
 952         }
 953         return 0;
 954 }
 955
 956 /*
 957  * copy a single inode. do all the required works, such as cloning
 958  * inode item, creating file extents and creating directory entries.
 959  */
 960 static int copy_single_inode(struct btrfs_trans_handle *trans,
 961                              struct btrfs_root *root, u64 objectid,
 962                              ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 963                              struct ext2_inode *ext2_inode,
 964                              int datacsum, int packing, int noxattr)
 965 {
 966         int ret;
 967         struct btrfs_key inode_key;
 968         struct btrfs_inode_item btrfs_inode;
 969
 970         if (ext2_inode->i_links_count == 0)
 971                 return 0;
 972
 973         copy_inode_item(&btrfs_inode, ext2_inode, ext2_fs->blocksize);
 974         if (!datacsum && S_ISREG(ext2_inode->i_mode)) {
 975                 u32 flags = btrfs_stack_inode_flags(&btrfs_inode) |
 976                             BTRFS_INODE_NODATASUM;
 977                 btrfs_set_stack_inode_flags(&btrfs_inode, flags);
 978         }
 979
 980         switch (ext2_inode->i_mode & S_IFMT) {
 981         case S_IFREG:
 982                 ret = create_file_extents(trans, root, objectid, &btrfs_inode,
 983                                         ext2_fs, ext2_ino, datacsum, packing);
 984                 break;
 985         case S_IFDIR:
 986                 ret = create_dir_entries(trans, root, objectid, &btrfs_inode,
 987                                          ext2_fs, ext2_ino);
 988                 break;
 989         case S_IFLNK:
 990                 ret = create_symbol_link(trans, root, objectid, &btrfs_inode,
 991                                          ext2_fs, ext2_ino, ext2_inode);
 992                 break;
 993         default:
 994                 ret = 0;
 995                 break;
 996         }
 997         if (ret)
 998                 return ret;
 999
1000         if (!noxattr) {
1001                 ret = copy_extended_attrs(trans, root, objectid, &btrfs_inode,
1002                                           ext2_fs, ext2_ino);
1003                 if (ret)
1004                         return ret;
1005         }
1006         inode_key.objectid = objectid;
1007         inode_key.offset = 0;
1008         btrfs_set_key_type(&inode_key, BTRFS_INODE_ITEM_KEY);
1009         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1010         return ret;
1011 }
1012
1013 static int copy_disk_extent(struct btrfs_root *root, u64 dst_bytenr,
1014                             u64 src_bytenr, u32 num_bytes)
1015 {
1016         int ret;
1017         char *buffer;
1018         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
1019
1020         buffer = malloc(num_bytes);
1021         if (!buffer)
1022                 return -ENOMEM;
1023         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, src_bytenr);
1024         if (ret != num_bytes)
1025                 goto fail;
1026         ret = pwrite(fs_devs->latest_bdev, buffer, num_bytes, dst_bytenr);
1027         if (ret != num_bytes)
1028                 goto fail;
1029         ret = 0;
1030 fail:
1031         free(buffer);
1032         if (ret > 0)
1033                 ret = -1;
1034         return ret;
1035 }
1036 /*
1037  * scan ext2's inode bitmap and copy all used inodes.
1038  */
1039 static int copy_inodes(struct btrfs_root *root, ext2_filsys ext2_fs,
1040                        int datacsum, int packing, int noxattr)
1041 {
1042         int ret;
1043         errcode_t err;
1044         ext2_inode_scan ext2_scan;
1045         struct ext2_inode ext2_inode;
1046         ext2_ino_t ext2_ino;
1047         u64 objectid;
1048         struct btrfs_trans_handle *trans;
1049
1050         trans = btrfs_start_transaction(root, 1);
1051         if (!trans)
1052                 return -ENOMEM;
1053         err = ext2fs_open_inode_scan(ext2_fs, 0, &ext2_scan);
1054         if (err) {
1055                 fprintf(stderr, "ext2fs_open_inode_scan: %s\n", error_message(err));
1056                 return -1;
1057         }
1058         while (!(err = ext2fs_get_next_inode(ext2_scan, &ext2_ino,
1059                                              &ext2_inode))) {
1060                 /* no more inodes */
1061                 if (ext2_ino == 0)
1062                         break;
1063                 /* skip special inode in ext2fs */
1064                 if (ext2_ino < EXT2_GOOD_OLD_FIRST_INO &&
1065                     ext2_ino != EXT2_ROOT_INO)
1066                         continue;
1067                 objectid = ext2_ino + INO_OFFSET;
1068                 ret = copy_single_inode(trans, root,
1069                                         objectid, ext2_fs, ext2_ino,
1070                                         &ext2_inode, datacsum, packing,
1071                                         noxattr);
1072                 if (ret)
1073                         return ret;
1074                 if (trans->blocks_used >= 4096) {
1075                         ret = btrfs_commit_transaction(trans, root);
1076                         BUG_ON(ret);
1077                         trans = btrfs_start_transaction(root, 1);
1078                         BUG_ON(!trans);
1079                 }
1080         }
1081         if (err) {
1082                 fprintf(stderr, "ext2fs_get_next_inode: %s\n", error_message(err));
1083                 return -1;
1084         }
1085         ret = btrfs_commit_transaction(trans, root);
1086         BUG_ON(ret);
1087
1088         return ret;
1089 }
1090
1091 /*
1092  * Construct a range of ext2fs image file.
1093  * scan block allocation bitmap, find all blocks used by the ext2fs
1094  * in this range and create file extents that point to these blocks.
1095  *
1096  * Note: Before calling the function, no file extent points to blocks
1097  *       in this range
1098  */
1099 static int create_image_file_range(struct btrfs_trans_handle *trans,
1100                                    struct btrfs_root *root, u64 objectid,
1101                                    struct btrfs_inode_item *inode,
1102                                    u64 start_byte, u64 end_byte,
1103                                    ext2_filsys ext2_fs)
1104 {
1105         u32 blocksize = ext2_fs->blocksize;
1106         u32 block = start_byte / blocksize;
1107         u32 last_block = (end_byte + blocksize - 1) / blocksize;
1108         int ret = 0;
1109         struct blk_iterate_data data = {
1110                 .trans          = trans,
1111                 .root           = root,
1112                 .inode          = inode,
1113                 .objectid       = objectid,
1114                 .first_block    = block,
1115                 .disk_block     = 0,
1116                 .num_blocks     = 0,
1117                 .boundary       = (u64)-1,
1118                 .checksum       = 0,
1119                 .errcode        = 0,
1120         };
1121         for (; start_byte < end_byte; block++, start_byte += blocksize) {
1122                 if (!ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
1123                         continue;
1124                 ret = block_iterate_proc(NULL, block, block, &data);
1125                 if (ret & BLOCK_ABORT) {
1126                         ret = data.errcode;
1127                         goto fail;
1128                 }
1129         }
1130         if (data.num_blocks > 0) {
1131                 ret = record_file_blocks(trans, root, objectid, inode,
1132                                          data.first_block, data.disk_block,
1133                                          data.num_blocks, 0);
1134                 if (ret)
1135                         goto fail;
1136                 data.first_block += data.num_blocks;
1137         }
1138         if (last_block > data.first_block) {
1139                 ret = record_file_blocks(trans, root, objectid, inode,
1140                                          data.first_block, 0, last_block -
1141                                          data.first_block, 0);
1142                 if (ret)
1143                         goto fail;
1144         }
1145 fail:
1146         return ret;
1147 }
1148 /*
1149  * Create the ext2fs image file.
1150  */
1151 static int create_ext2_image(struct btrfs_root *root, ext2_filsys ext2_fs,
1152                              const char *name)
1153 {
1154         int ret;
1155         struct btrfs_key key;
1156         struct btrfs_key location;
1157         struct btrfs_path path;
1158         struct btrfs_inode_item btrfs_inode;
1159         struct btrfs_inode_item *inode_item;
1160         struct extent_buffer *leaf;
1161         struct btrfs_fs_info *fs_info = root->fs_info;
1162         struct btrfs_root *extent_root = fs_info->extent_root;
1163         struct btrfs_trans_handle *trans;
1164         struct btrfs_extent_item *ei;
1165         struct btrfs_extent_inline_ref *iref;
1166         struct btrfs_extent_data_ref *dref;
1167         u64 bytenr;
1168         u64 num_bytes;
1169         u64 objectid;
1170         u64 last_byte;
1171         u64 first_free;
1172         u64 total_bytes;
1173         u32 sectorsize = root->sectorsize;
1174
1175         total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
1176         first_free =  BTRFS_SUPER_INFO_OFFSET + sectorsize * 2 - 1;
1177         first_free &= ~((u64)sectorsize - 1);
1178
1179         memset(&btrfs_inode, 0, sizeof(btrfs_inode));
1180         btrfs_set_stack_inode_generation(&btrfs_inode, 1);
1181         btrfs_set_stack_inode_size(&btrfs_inode, total_bytes);
1182         btrfs_set_stack_inode_nlink(&btrfs_inode, 1);
1183         btrfs_set_stack_inode_nbytes(&btrfs_inode, 0);
1184         btrfs_set_stack_inode_mode(&btrfs_inode, S_IFREG | 0400);
1185         btrfs_set_stack_inode_flags(&btrfs_inode, BTRFS_INODE_NODATASUM |
1186                                     BTRFS_INODE_READONLY);
1187         btrfs_init_path(&path);
1188         trans = btrfs_start_transaction(root, 1);
1189         BUG_ON(!trans);
1190
1191         objectid = btrfs_root_dirid(&root->root_item);
1192         ret = btrfs_find_free_objectid(trans, root, objectid, &objectid);
1193         if (ret)
1194                 goto fail;
1195
1196         /*
1197          * copy blocks covered by extent #0 to new positions. extent #0 is
1198          * special, we can't rely on relocate_extents_range to relocate it.
1199          */
1200         for (last_byte = 0; last_byte < first_free; last_byte += sectorsize) {
1201                 ret = custom_alloc_extent(root, sectorsize, 0, &key);
1202                 if (ret)
1203                         goto fail;
1204                 ret = copy_disk_extent(root, key.objectid, last_byte,
1205                                        sectorsize);
1206                 if (ret)
1207                         goto fail;
1208                 ret = btrfs_record_file_extent(trans, root, objectid,
1209                                                &btrfs_inode, last_byte,
1210                                                key.objectid, sectorsize);
1211                 if (ret)
1212                         goto fail;
1213         }
1214
1215         while(1) {
1216                 key.objectid = last_byte;
1217                 key.offset = 0;
1218                 btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
1219                 ret = btrfs_search_slot(trans, fs_info->extent_root,
1220                                         &key, &path, 0, 0);
1221                 if (ret < 0)
1222                         goto fail;
1223 next:
1224                 leaf = path.nodes[0];
1225                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1226                         ret = btrfs_next_leaf(extent_root, &path);
1227                         if (ret < 0)
1228                                 goto fail;
1229                         if (ret > 0)
1230                                 break;
1231                         leaf = path.nodes[0];
1232                 }
1233                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1234                 if (last_byte > key.objectid ||
1235                     key.type != BTRFS_EXTENT_ITEM_KEY) {
1236                         path.slots[0]++;
1237                         goto next;
1238                 }
1239
1240                 bytenr = key.objectid;
1241                 num_bytes = key.offset;
1242                 ei = btrfs_item_ptr(leaf, path.slots[0],
1243                                     struct btrfs_extent_item);
1244                 if (!(btrfs_extent_flags(leaf, ei) & BTRFS_EXTENT_FLAG_DATA)) {
1245                         path.slots[0]++;
1246                         goto next;
1247                 }
1248
1249                 BUG_ON(btrfs_item_size_nr(leaf, path.slots[0]) != sizeof(*ei) +
1250                        btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY));
1251
1252                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
1253                 key.type = btrfs_extent_inline_ref_type(leaf, iref);
1254                 BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
1255                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1256                 if (btrfs_extent_data_ref_root(leaf, dref) !=
1257                     BTRFS_FS_TREE_OBJECTID) {
1258                         path.slots[0]++;
1259                         goto next;
1260                 }
1261
1262                 if (bytenr > last_byte) {
1263                         ret = create_image_file_range(trans, root, objectid,
1264                                                       &btrfs_inode, last_byte,
1265                                                       bytenr, ext2_fs);
1266                         if (ret)
1267                                 goto fail;
1268                 }
1269                 ret = btrfs_record_file_extent(trans, root, objectid,
1270                                                &btrfs_inode, bytenr, bytenr,
1271                                                num_bytes);
1272                 if (ret)
1273                         goto fail;
1274                 last_byte = bytenr + num_bytes;
1275                 btrfs_release_path(&path);
1276
1277                 if (trans->blocks_used >= 4096) {
1278                         ret = btrfs_commit_transaction(trans, root);
1279                         BUG_ON(ret);
1280                         trans = btrfs_start_transaction(root, 1);
1281                         BUG_ON(!trans);
1282                 }
1283         }
1284         btrfs_release_path(&path);
1285         if (total_bytes > last_byte) {
1286                 ret = create_image_file_range(trans, root, objectid,
1287                                               &btrfs_inode, last_byte,
1288                                               total_bytes, ext2_fs);
1289                 if (ret)
1290                         goto fail;
1291         }
1292
1293         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1294         if (ret)
1295                 goto fail;
1296
1297         location.objectid = objectid;
1298         location.offset = 0;
1299         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1300         ret = btrfs_insert_dir_item(trans, root, name, strlen(name),
1301                                     btrfs_root_dirid(&root->root_item),
1302                                     &location, EXT2_FT_REG_FILE, objectid);
1303         if (ret)
1304                 goto fail;
1305         ret = btrfs_insert_inode_ref(trans, root, name, strlen(name),
1306                                      objectid,
1307                                      btrfs_root_dirid(&root->root_item),
1308                                      objectid);
1309         if (ret)
1310                 goto fail;
1311         location.objectid = btrfs_root_dirid(&root->root_item);
1312         location.offset = 0;
1313         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1314         ret = btrfs_lookup_inode(trans, root, &path, &location, 1);
1315         if (ret)
1316                 goto fail;
1317         leaf = path.nodes[0];
1318         inode_item = btrfs_item_ptr(leaf, path.slots[0],
1319                                     struct btrfs_inode_item);
1320         btrfs_set_inode_size(leaf, inode_item, strlen(name) * 2 +
1321                              btrfs_inode_size(leaf, inode_item));
1322         btrfs_mark_buffer_dirty(leaf);
1323         btrfs_release_path(&path);
1324         ret = btrfs_commit_transaction(trans, root);
1325         BUG_ON(ret);
1326 fail:
1327         btrfs_release_path(&path);
1328         return ret;
1329 }
1330
1331 static struct btrfs_root * link_subvol(struct btrfs_root *root,
1332                 const char *base, u64 root_objectid)
1333 {
1334         struct btrfs_trans_handle *trans;
1335         struct btrfs_fs_info *fs_info = root->fs_info;
1336         struct btrfs_root *tree_root = fs_info->tree_root;
1337         struct btrfs_root *new_root = NULL;
1338         struct btrfs_path *path;
1339         struct btrfs_inode_item *inode_item;
1340         struct extent_buffer *leaf;
1341         struct btrfs_key key;
1342         u64 dirid = btrfs_root_dirid(&root->root_item);
1343         u64 index = 2;
1344         char buf[BTRFS_NAME_LEN + 1]; /* for snprintf null */
1345         int len;
1346         int i;
1347         int ret;
1348
1349         len = strlen(base);
1350         if (len < 1 || len > BTRFS_NAME_LEN)
1351                 return NULL;
1352
1353         path = btrfs_alloc_path();
1354         BUG_ON(!path);
1355
1356         key.objectid = dirid;
1357         key.type = BTRFS_DIR_INDEX_KEY;
1358         key.offset = (u64)-1;
1359
1360         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1361         BUG_ON(ret <= 0);
1362
1363         if (path->slots[0] > 0) {
1364                 path->slots[0]--;
1365                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1366                 if (key.objectid == dirid && key.type == BTRFS_DIR_INDEX_KEY)
1367                         index = key.offset + 1;
1368         }
1369         btrfs_release_path(path);
1370
1371         trans = btrfs_start_transaction(root, 1);
1372         BUG_ON(!trans);
1373
1374         key.objectid = dirid;
1375         key.offset = 0;
1376         key.type =  BTRFS_INODE_ITEM_KEY;
1377
1378         ret = btrfs_lookup_inode(trans, root, path, &key, 1);
1379         BUG_ON(ret);
1380         leaf = path->nodes[0];
1381         inode_item = btrfs_item_ptr(leaf, path->slots[0],
1382                                     struct btrfs_inode_item);
1383
1384         key.objectid = root_objectid;
1385         key.offset = (u64)-1;
1386         key.type = BTRFS_ROOT_ITEM_KEY;
1387
1388         memcpy(buf, base, len);
1389         for (i = 0; i < 1024; i++) {
1390                 ret = btrfs_insert_dir_item(trans, root, buf, len,
1391                                             dirid, &key, BTRFS_FT_DIR, index);
1392                 if (ret != -EEXIST)
1393                         break;
1394                 len = snprintf(buf, ARRAY_SIZE(buf), "%s%d", base, i);
1395                 if (len < 1 || len > BTRFS_NAME_LEN) {
1396                         ret = -EINVAL;
1397                         break;
1398                 }
1399         }
1400         if (ret)
1401                 goto fail;
1402
1403         btrfs_set_inode_size(leaf, inode_item, len * 2 +
1404                              btrfs_inode_size(leaf, inode_item));
1405         btrfs_mark_buffer_dirty(leaf);
1406         btrfs_release_path(path);
1407
1408         /* add the backref first */
1409         ret = btrfs_add_root_ref(trans, tree_root, root_objectid,
1410                                  BTRFS_ROOT_BACKREF_KEY,
1411                                  root->root_key.objectid,
1412                                  dirid, index, buf, len);
1413         BUG_ON(ret);
1414
1415         /* now add the forward ref */
1416         ret = btrfs_add_root_ref(trans, tree_root, root->root_key.objectid,
1417                                  BTRFS_ROOT_REF_KEY, root_objectid,
1418                                  dirid, index, buf, len);
1419
1420         ret = btrfs_commit_transaction(trans, root);
1421         BUG_ON(ret);
1422
1423         new_root = btrfs_read_fs_root(fs_info, &key);
1424         if (IS_ERR(new_root))
1425                 new_root = NULL;
1426 fail:
1427         btrfs_free_path(path);
1428         return new_root;
1429 }
1430
1431 static int create_chunk_mapping(struct btrfs_trans_handle *trans,
1432                                 struct btrfs_root *root)
1433 {
1434         struct btrfs_fs_info *info = root->fs_info;
1435         struct btrfs_root *chunk_root = info->chunk_root;
1436         struct btrfs_root *extent_root = info->extent_root;
1437         struct btrfs_device *device;
1438         struct btrfs_block_group_cache *cache;
1439         struct btrfs_dev_extent *extent;
1440         struct extent_buffer *leaf;
1441         struct btrfs_chunk chunk;
1442         struct btrfs_key key;
1443         struct btrfs_path path;
1444         u64 cur_start;
1445         u64 total_bytes;
1446         u64 chunk_objectid;
1447         int ret;
1448
1449         btrfs_init_path(&path);
1450
1451         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
1452         chunk_objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
1453
1454         BUG_ON(list_empty(&info->fs_devices->devices));
1455         device = list_entry(info->fs_devices->devices.next,
1456                             struct btrfs_device, dev_list);
1457         BUG_ON(device->devid != info->fs_devices->latest_devid);
1458
1459         /* delete device extent created by make_btrfs */
1460         key.objectid = device->devid;
1461         key.offset = 0;
1462         key.type = BTRFS_DEV_EXTENT_KEY;
1463         ret = btrfs_search_slot(trans, device->dev_root, &key, &path, -1, 1);
1464         if (ret < 0)
1465                 goto err;
1466
1467         BUG_ON(ret > 0);
1468         ret = btrfs_del_item(trans, device->dev_root, &path);
1469         if (ret)
1470                 goto err;
1471         btrfs_release_path(&path);
1472
1473         /* delete chunk item created by make_btrfs */
1474         key.objectid = chunk_objectid;
1475         key.offset = 0;
1476         key.type = BTRFS_CHUNK_ITEM_KEY;
1477         ret = btrfs_search_slot(trans, chunk_root, &key, &path, -1, 1);
1478         if (ret < 0)
1479                 goto err;
1480
1481         BUG_ON(ret > 0);
1482         ret = btrfs_del_item(trans, chunk_root, &path);
1483         if (ret)
1484                 goto err;
1485         btrfs_release_path(&path);
1486
1487         /* for each block group, create device extent and chunk item */
1488         cur_start = 0;
1489         while (cur_start < total_bytes) {
1490                 cache = btrfs_lookup_block_group(root->fs_info, cur_start);
1491                 BUG_ON(!cache);
1492
1493                 /* insert device extent */
1494                 key.objectid = device->devid;
1495                 key.offset = cache->key.objectid;
1496                 key.type = BTRFS_DEV_EXTENT_KEY;
1497                 ret = btrfs_insert_empty_item(trans, device->dev_root, &path,
1498                                               &key, sizeof(*extent));
1499                 if (ret)
1500                         goto err;
1501
1502                 leaf = path.nodes[0];
1503                 extent = btrfs_item_ptr(leaf, path.slots[0],
1504                                         struct btrfs_dev_extent);
1505
1506                 btrfs_set_dev_extent_chunk_tree(leaf, extent,
1507                                                 chunk_root->root_key.objectid);
1508                 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
1509                                                     chunk_objectid);
1510                 btrfs_set_dev_extent_chunk_offset(leaf, extent,
1511                                                   cache->key.objectid);
1512                 btrfs_set_dev_extent_length(leaf, extent, cache->key.offset);
1513                 write_extent_buffer(leaf, root->fs_info->chunk_tree_uuid,
1514                     (unsigned long)btrfs_dev_extent_chunk_tree_uuid(extent),
1515                     BTRFS_UUID_SIZE);
1516                 btrfs_mark_buffer_dirty(leaf);
1517                 btrfs_release_path(&path);
1518
1519                 /* insert chunk item */
1520                 btrfs_set_stack_chunk_length(&chunk, cache->key.offset);
1521                 btrfs_set_stack_chunk_owner(&chunk,
1522                                             extent_root->root_key.objectid);
1523                 btrfs_set_stack_chunk_stripe_len(&chunk, STRIPE_LEN);
1524                 btrfs_set_stack_chunk_type(&chunk, cache->flags);
1525                 btrfs_set_stack_chunk_io_align(&chunk, device->io_align);
1526                 btrfs_set_stack_chunk_io_width(&chunk, device->io_width);
1527                 btrfs_set_stack_chunk_sector_size(&chunk, device->sector_size);
1528                 btrfs_set_stack_chunk_num_stripes(&chunk, 1);
1529                 btrfs_set_stack_chunk_sub_stripes(&chunk, 0);
1530                 btrfs_set_stack_stripe_devid(&chunk.stripe, device->devid);
1531                 btrfs_set_stack_stripe_offset(&chunk.stripe,
1532                                               cache->key.objectid);
1533                 memcpy(&chunk.stripe.dev_uuid, device->uuid, BTRFS_UUID_SIZE);
1534
1535                 key.objectid = chunk_objectid;
1536                 key.offset = cache->key.objectid;
1537                 key.type = BTRFS_CHUNK_ITEM_KEY;
1538
1539                 ret = btrfs_insert_item(trans, chunk_root, &key, &chunk,
1540                                         btrfs_chunk_item_size(1));
1541                 if (ret)
1542                         goto err;
1543
1544                 cur_start = cache->key.objectid + cache->key.offset;
1545         }
1546
1547         device->bytes_used = total_bytes;
1548         ret = btrfs_update_device(trans, device);
1549 err:
1550         btrfs_release_path(&path);
1551         return ret;
1552 }
1553
1554 static int create_subvol(struct btrfs_trans_handle *trans,
1555                          struct btrfs_root *root, u64 root_objectid)
1556 {
1557         struct extent_buffer *tmp;
1558         struct btrfs_root *new_root;
1559         struct btrfs_key key;
1560         struct btrfs_root_item root_item;
1561         int ret;
1562
1563         ret = btrfs_copy_root(trans, root, root->node, &tmp,
1564                               root_objectid);
1565         BUG_ON(ret);
1566
1567         memcpy(&root_item, &root->root_item, sizeof(root_item));
1568         btrfs_set_root_bytenr(&root_item, tmp->start);
1569         btrfs_set_root_level(&root_item, btrfs_header_level(tmp));
1570         btrfs_set_root_generation(&root_item, trans->transid);
1571         free_extent_buffer(tmp);
1572
1573         key.objectid = root_objectid;
1574         key.type = BTRFS_ROOT_ITEM_KEY;
1575         key.offset = trans->transid;
1576         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1577                                 &key, &root_item);
1578
1579         key.offset = (u64)-1;
1580         new_root = btrfs_read_fs_root(root->fs_info, &key);
1581         BUG_ON(!new_root || IS_ERR(new_root));
1582
1583         ret = btrfs_make_root_dir(trans, new_root, BTRFS_FIRST_FREE_OBJECTID);
1584         BUG_ON(ret);
1585
1586         return 0;
1587 }
1588
1589 static int init_btrfs(struct btrfs_root *root)
1590 {
1591         int ret;
1592         struct btrfs_key location;
1593         struct btrfs_trans_handle *trans;
1594         struct btrfs_fs_info *fs_info = root->fs_info;
1595         struct extent_buffer *tmp;
1596
1597         trans = btrfs_start_transaction(root, 1);
1598         BUG_ON(!trans);
1599         ret = btrfs_make_block_groups(trans, root);
1600         if (ret)
1601                 goto err;
1602         ret = btrfs_fix_block_accounting(trans, root);
1603         if (ret)
1604                 goto err;
1605         ret = create_chunk_mapping(trans, root);
1606         if (ret)
1607                 goto err;
1608         ret = btrfs_make_root_dir(trans, fs_info->tree_root,
1609                                   BTRFS_ROOT_TREE_DIR_OBJECTID);
1610         if (ret)
1611                 goto err;
1612         memcpy(&location, &root->root_key, sizeof(location));
1613         location.offset = (u64)-1;
1614         ret = btrfs_insert_dir_item(trans, fs_info->tree_root, "default", 7,
1615                                 btrfs_super_root_dir(fs_info->super_copy),
1616                                 &location, BTRFS_FT_DIR, 0);
1617         if (ret)
1618                 goto err;
1619         ret = btrfs_insert_inode_ref(trans, fs_info->tree_root, "default", 7,
1620                                 location.objectid,
1621                                 btrfs_super_root_dir(fs_info->super_copy), 0);
1622         if (ret)
1623                 goto err;
1624         btrfs_set_root_dirid(&fs_info->fs_root->root_item,
1625                              BTRFS_FIRST_FREE_OBJECTID);
1626
1627         /* subvol for ext2 image file */
1628         ret = create_subvol(trans, root, EXT2_IMAGE_SUBVOL_OBJECTID);
1629         BUG_ON(ret);
1630         /* subvol for data relocation */
1631         ret = create_subvol(trans, root, BTRFS_DATA_RELOC_TREE_OBJECTID);
1632         BUG_ON(ret);
1633
1634         extent_buffer_get(fs_info->csum_root->node);
1635         ret = __btrfs_cow_block(trans, fs_info->csum_root,
1636                                 fs_info->csum_root->node, NULL, 0, &tmp, 0, 0);
1637         BUG_ON(ret);
1638         free_extent_buffer(tmp);
1639
1640         ret = btrfs_commit_transaction(trans, root);
1641         BUG_ON(ret);
1642 err:
1643         return ret;
1644 }
1645
1646 /*
1647  * Migrate super block to it's default position and zero 0 ~ 16k
1648  */
1649 static int migrate_super_block(int fd, u64 old_bytenr, u32 sectorsize)
1650 {
1651         int ret;
1652         struct extent_buffer *buf;
1653         struct btrfs_super_block *super;
1654         u32 len;
1655         u32 bytenr;
1656
1657         BUG_ON(sectorsize < sizeof(*super));
1658         buf = malloc(sizeof(*buf) + sectorsize);
1659         if (!buf)
1660                 return -ENOMEM;
1661
1662         buf->len = sectorsize;
1663         ret = pread(fd, buf->data, sectorsize, old_bytenr);
1664         if (ret != sectorsize)
1665                 goto fail;
1666
1667         super = (struct btrfs_super_block *)buf->data;
1668         BUG_ON(btrfs_super_bytenr(super) != old_bytenr);
1669         btrfs_set_super_bytenr(super, BTRFS_SUPER_INFO_OFFSET);
1670
1671         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1672         ret = pwrite(fd, buf->data, sectorsize, BTRFS_SUPER_INFO_OFFSET);
1673         if (ret != sectorsize)
1674                 goto fail;
1675
1676         ret = fsync(fd);
1677         if (ret)
1678                 goto fail;
1679
1680         memset(buf->data, 0, sectorsize);
1681         for (bytenr = 0; bytenr < BTRFS_SUPER_INFO_OFFSET; ) {
1682                 len = BTRFS_SUPER_INFO_OFFSET - bytenr;
1683                 if (len > sectorsize)
1684                         len = sectorsize;
1685                 ret = pwrite(fd, buf->data, len, bytenr);
1686                 if (ret != len) {
1687                         fprintf(stderr, "unable to zero fill device\n");
1688                         break;
1689                 }
1690                 bytenr += len;
1691         }
1692         ret = 0;
1693         fsync(fd);
1694 fail:
1695         free(buf);
1696         if (ret > 0)
1697                 ret = -1;
1698         return ret;
1699 }
1700
1701 static int prepare_system_chunk_sb(struct btrfs_super_block *super)
1702 {
1703         struct btrfs_chunk *chunk;
1704         struct btrfs_disk_key *key;
1705         u32 sectorsize = btrfs_super_sectorsize(super);
1706
1707         key = (struct btrfs_disk_key *)(super->sys_chunk_array);
1708         chunk = (struct btrfs_chunk *)(super->sys_chunk_array +
1709                                        sizeof(struct btrfs_disk_key));
1710
1711         btrfs_set_disk_key_objectid(key, BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1712         btrfs_set_disk_key_type(key, BTRFS_CHUNK_ITEM_KEY);
1713         btrfs_set_disk_key_offset(key, 0);
1714
1715         btrfs_set_stack_chunk_length(chunk, btrfs_super_total_bytes(super));
1716         btrfs_set_stack_chunk_owner(chunk, BTRFS_EXTENT_TREE_OBJECTID);
1717         btrfs_set_stack_chunk_stripe_len(chunk, 64 * 1024);
1718         btrfs_set_stack_chunk_type(chunk, BTRFS_BLOCK_GROUP_SYSTEM);
1719         btrfs_set_stack_chunk_io_align(chunk, sectorsize);
1720         btrfs_set_stack_chunk_io_width(chunk, sectorsize);
1721         btrfs_set_stack_chunk_sector_size(chunk, sectorsize);
1722         btrfs_set_stack_chunk_num_stripes(chunk, 1);
1723         btrfs_set_stack_chunk_sub_stripes(chunk, 0);
1724         chunk->stripe.devid = super->dev_item.devid;
1725         btrfs_set_stack_stripe_offset(&chunk->stripe, 0);
1726         memcpy(chunk->stripe.dev_uuid, super->dev_item.uuid, BTRFS_UUID_SIZE);
1727         btrfs_set_super_sys_array_size(super, sizeof(*key) + sizeof(*chunk));
1728         return 0;
1729 }
1730
1731 static int prepare_system_chunk(int fd, u64 sb_bytenr, u32 sectorsize)
1732 {
1733         int ret;
1734         struct extent_buffer *buf;
1735         struct btrfs_super_block *super;
1736
1737         BUG_ON(sectorsize < sizeof(*super));
1738         buf = malloc(sizeof(*buf) + sectorsize);
1739         if (!buf)
1740                 return -ENOMEM;
1741
1742         buf->len = sectorsize;
1743         ret = pread(fd, buf->data, sectorsize, sb_bytenr);
1744         if (ret != sectorsize)
1745                 goto fail;
1746
1747         super = (struct btrfs_super_block *)buf->data;
1748         BUG_ON(btrfs_super_bytenr(super) != sb_bytenr);
1749         BUG_ON(btrfs_super_num_devices(super) != 1);
1750
1751         ret = prepare_system_chunk_sb(super);
1752         if (ret)
1753                 goto fail;
1754
1755         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1756         ret = pwrite(fd, buf->data, sectorsize, sb_bytenr);
1757         if (ret != sectorsize)
1758                 goto fail;
1759
1760         ret = 0;
1761 fail:
1762         free(buf);
1763         if (ret > 0)
1764                 ret = -1;
1765         return ret;
1766 }
1767
1768 static int relocate_one_reference(struct btrfs_trans_handle *trans,
1769                                   struct btrfs_root *root,
1770                                   u64 extent_start, u64 extent_size,
1771                                   struct btrfs_key *extent_key,
1772                                   struct extent_io_tree *reloc_tree)
1773 {
1774         struct extent_buffer *leaf;
1775         struct btrfs_file_extent_item *fi;
1776         struct btrfs_key key;
1777         struct btrfs_path path;
1778         struct btrfs_inode_item inode;
1779         struct blk_iterate_data data;
1780         u64 bytenr;
1781         u64 num_bytes;
1782         u64 cur_offset;
1783         u64 new_pos;
1784         u64 nbytes;
1785         u64 sector_end;
1786         u32 sectorsize = root->sectorsize;
1787         unsigned long ptr;
1788         int datacsum;
1789         int fd;
1790         int ret;
1791
1792         btrfs_init_path(&path);
1793         ret = btrfs_search_slot(trans, root, extent_key, &path, -1, 1);
1794         if (ret)
1795                 goto fail;
1796
1797         leaf = path.nodes[0];
1798         fi = btrfs_item_ptr(leaf, path.slots[0],
1799                             struct btrfs_file_extent_item);
1800         BUG_ON(btrfs_file_extent_offset(leaf, fi) > 0);
1801         if (extent_start != btrfs_file_extent_disk_bytenr(leaf, fi) ||
1802             extent_size != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1803                 ret = 1;
1804                 goto fail;
1805         }
1806
1807         bytenr = extent_start + btrfs_file_extent_offset(leaf, fi);
1808         num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
1809
1810         ret = btrfs_del_item(trans, root, &path);
1811         if (ret)
1812                 goto fail;
1813
1814         ret = btrfs_free_extent(trans, root, extent_start, extent_size, 0,
1815                                 root->root_key.objectid,
1816                                 extent_key->objectid, extent_key->offset);
1817         if (ret)
1818                 goto fail;
1819
1820         btrfs_release_path(&path);
1821
1822         key.objectid = extent_key->objectid;
1823         key.offset = 0;
1824         key.type =  BTRFS_INODE_ITEM_KEY;
1825         ret = btrfs_lookup_inode(trans, root, &path, &key, 0);
1826         if (ret)
1827                 goto fail;
1828
1829         leaf = path.nodes[0];
1830         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1831         read_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1832         btrfs_release_path(&path);
1833
1834         BUG_ON(num_bytes & (sectorsize - 1));
1835         nbytes = btrfs_stack_inode_nbytes(&inode) - num_bytes;
1836         btrfs_set_stack_inode_nbytes(&inode, nbytes);
1837         datacsum = !(btrfs_stack_inode_flags(&inode) & BTRFS_INODE_NODATASUM);
1838
1839         data = (struct blk_iterate_data) {
1840                 .trans          = trans,
1841                 .root           = root,
1842                 .inode          = &inode,
1843                 .objectid       = extent_key->objectid,
1844                 .first_block    = extent_key->offset / sectorsize,
1845                 .disk_block     = 0,
1846                 .num_blocks     = 0,
1847                 .boundary       = (u64)-1,
1848                 .checksum       = datacsum,
1849                 .errcode        = 0,
1850         };
1851
1852         cur_offset = extent_key->offset;
1853         while (num_bytes > 0) {
1854                 sector_end = bytenr + sectorsize - 1;
1855                 if (test_range_bit(reloc_tree, bytenr, sector_end,
1856                                    EXTENT_LOCKED, 1)) {
1857                         ret = get_state_private(reloc_tree, bytenr, &new_pos);
1858                         BUG_ON(ret);
1859                 } else {
1860                         ret = custom_alloc_extent(root, sectorsize, 0, &key);
1861                         if (ret)
1862                                 goto fail;
1863                         new_pos = key.objectid;
1864
1865                         if (cur_offset == extent_key->offset) {
1866                                 fd = root->fs_info->fs_devices->latest_bdev;
1867                                 readahead(fd, bytenr, num_bytes);
1868                         }
1869                         ret = copy_disk_extent(root, new_pos, bytenr,
1870                                                sectorsize);
1871                         if (ret)
1872                                 goto fail;
1873                         ret = set_extent_bits(reloc_tree, bytenr, sector_end,
1874                                               EXTENT_LOCKED, GFP_NOFS);
1875                         BUG_ON(ret);
1876                         ret = set_state_private(reloc_tree, bytenr, new_pos);
1877                         BUG_ON(ret);
1878                 }
1879
1880                 ret = block_iterate_proc(NULL, new_pos / sectorsize,
1881                                          cur_offset / sectorsize, &data);
1882                 if (ret & BLOCK_ABORT) {
1883                         ret = data.errcode;
1884                         goto fail;
1885                 }
1886
1887                 cur_offset += sectorsize;
1888                 bytenr += sectorsize;
1889                 num_bytes -= sectorsize;
1890         }
1891
1892         if (data.num_blocks > 0) {
1893                 ret = record_file_blocks(trans, root,
1894                                          extent_key->objectid, &inode,
1895                                          data.first_block, data.disk_block,
1896                                          data.num_blocks, datacsum);
1897                 if (ret)
1898                         goto fail;
1899         }
1900
1901         key.objectid = extent_key->objectid;
1902         key.offset = 0;
1903         key.type =  BTRFS_INODE_ITEM_KEY;
1904         ret = btrfs_lookup_inode(trans, root, &path, &key, 1);
1905         if (ret)
1906                 goto fail;
1907
1908         leaf = path.nodes[0];
1909         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1910         write_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1911         btrfs_mark_buffer_dirty(leaf);
1912         btrfs_release_path(&path);
1913
1914 fail:
1915         btrfs_release_path(&path);
1916         return ret;
1917 }
1918
1919 static int relocate_extents_range(struct btrfs_root *fs_root,
1920                                   struct btrfs_root *ext2_root,
1921                                   u64 start_byte, u64 end_byte)
1922 {
1923         struct btrfs_fs_info *info = fs_root->fs_info;
1924         struct btrfs_root *extent_root = info->extent_root;
1925         struct btrfs_root *cur_root = NULL;
1926         struct btrfs_trans_handle *trans;
1927         struct btrfs_extent_data_ref *dref;
1928         struct btrfs_extent_inline_ref *iref;
1929         struct btrfs_extent_item *ei;
1930         struct extent_buffer *leaf;
1931         struct btrfs_key key;
1932         struct btrfs_key extent_key;
1933         struct btrfs_path path;
1934         struct extent_io_tree reloc_tree;
1935         unsigned long ptr;
1936         unsigned long end;
1937         u64 cur_byte;
1938         u64 num_bytes;
1939         u64 ref_root;
1940         u64 num_extents;
1941         int pass = 0;
1942         int ret;
1943
1944         btrfs_init_path(&path);
1945         extent_io_tree_init(&reloc_tree);
1946
1947         key.objectid = start_byte;
1948         key.offset = 0;
1949         key.type = BTRFS_EXTENT_ITEM_KEY;
1950         ret = btrfs_search_slot(NULL, extent_root, &key, &path, 0, 0);
1951         if (ret < 0)
1952                 goto fail;
1953         if (ret > 0) {
1954                 ret = btrfs_previous_item(extent_root, &path, 0,
1955                                           BTRFS_EXTENT_ITEM_KEY);
1956                 if (ret < 0)
1957                         goto fail;
1958                 if (ret == 0) {
1959                         leaf = path.nodes[0];
1960                         btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1961                         if (key.objectid + key.offset > start_byte)
1962                                 start_byte = key.objectid;
1963                 }
1964         }
1965         btrfs_release_path(&path);
1966 again:
1967         cur_root = (pass % 2 == 0) ? ext2_root : fs_root;
1968         num_extents = 0;
1969
1970         trans = btrfs_start_transaction(cur_root, 1);
1971         BUG_ON(!trans);
1972
1973         cur_byte = start_byte;
1974         while (1) {
1975                 key.objectid = cur_byte;
1976                 key.offset = 0;
1977                 key.type = BTRFS_EXTENT_ITEM_KEY;
1978                 ret = btrfs_search_slot(trans, extent_root,
1979                                         &key, &path, 0, 0);
1980                 if (ret < 0)
1981                         goto fail;
1982 next:
1983                 leaf = path.nodes[0];
1984                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1985                         ret = btrfs_next_leaf(extent_root, &path);
1986                         if (ret < 0)
1987                                 goto fail;
1988                         if (ret > 0)
1989                                 break;
1990                         leaf = path.nodes[0];
1991                 }
1992
1993                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1994                 if (key.objectid < cur_byte ||
1995                     key.type != BTRFS_EXTENT_ITEM_KEY) {
1996                         path.slots[0]++;
1997                         goto next;
1998                 }
1999                 if (key.objectid >= end_byte)
2000                         break;
2001
2002                 num_extents++;
2003
2004                 cur_byte = key.objectid;
2005                 num_bytes = key.offset;
2006                 ei = btrfs_item_ptr(leaf, path.slots[0],
2007                                     struct btrfs_extent_item);
2008                 BUG_ON(!(btrfs_extent_flags(leaf, ei) &
2009                          BTRFS_EXTENT_FLAG_DATA));
2010
2011                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2012                 end = ptr + btrfs_item_size_nr(leaf, path.slots[0]);
2013
2014                 ptr += sizeof(struct btrfs_extent_item);
2015
2016                 while (ptr < end) {
2017                         iref = (struct btrfs_extent_inline_ref *)ptr;
2018                         key.type = btrfs_extent_inline_ref_type(leaf, iref);
2019                         BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
2020                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2021                         ref_root = btrfs_extent_data_ref_root(leaf, dref);
2022                         extent_key.objectid =
2023                                 btrfs_extent_data_ref_objectid(leaf, dref);
2024                         extent_key.offset =
2025                                 btrfs_extent_data_ref_offset(leaf, dref);
2026                         extent_key.type = BTRFS_EXTENT_DATA_KEY;
2027                         BUG_ON(btrfs_extent_data_ref_count(leaf, dref) != 1);
2028
2029                         if (ref_root == cur_root->root_key.objectid)
2030                                 break;
2031
2032                         ptr += btrfs_extent_inline_ref_size(key.type);
2033                 }
2034
2035                 if (ptr >= end) {
2036                         path.slots[0]++;
2037                         goto next;
2038                 }
2039
2040                 ret = relocate_one_reference(trans, cur_root, cur_byte,
2041                                              num_bytes, &extent_key,
2042                                              &reloc_tree);
2043                 if (ret < 0)
2044                         goto fail;
2045
2046                 cur_byte += num_bytes;
2047                 btrfs_release_path(&path);
2048
2049                 if (trans->blocks_used >= 4096) {
2050                         ret = btrfs_commit_transaction(trans, cur_root);
2051                         BUG_ON(ret);
2052                         trans = btrfs_start_transaction(cur_root, 1);
2053                         BUG_ON(!trans);
2054                 }
2055         }
2056         btrfs_release_path(&path);
2057
2058         ret = btrfs_commit_transaction(trans, cur_root);
2059         BUG_ON(ret);
2060
2061         if (num_extents > 0 && pass++ < 16)
2062                 goto again;
2063
2064         ret = (num_extents > 0) ? -1 : 0;
2065 fail:
2066         btrfs_release_path(&path);
2067         extent_io_tree_cleanup(&reloc_tree);
2068         return ret;
2069 }
2070
2071 /*
2072  * relocate data in system chunk
2073  */
2074 static int cleanup_sys_chunk(struct btrfs_root *fs_root,
2075                              struct btrfs_root *ext2_root)
2076 {
2077         struct btrfs_block_group_cache *cache;
2078         int i, ret = 0;
2079         u64 offset = 0;
2080         u64 end_byte;
2081
2082         while(1) {
2083                 cache = btrfs_lookup_block_group(fs_root->fs_info, offset);
2084                 if (!cache)
2085                         break;
2086
2087                 end_byte = cache->key.objectid + cache->key.offset;
2088                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2089                         ret = relocate_extents_range(fs_root, ext2_root,
2090                                                      cache->key.objectid,
2091                                                      end_byte);
2092                         if (ret)
2093                                 goto fail;
2094                 }
2095                 offset = end_byte;
2096         }
2097         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2098                 offset = btrfs_sb_offset(i);
2099                 offset &= ~((u64)STRIPE_LEN - 1);
2100
2101                 ret = relocate_extents_range(fs_root, ext2_root,
2102                                              offset, offset + STRIPE_LEN);
2103                 if (ret)
2104                         goto fail;
2105         }
2106         ret = 0;
2107 fail:
2108         return ret;
2109 }
2110
2111 static int fixup_chunk_mapping(struct btrfs_root *root)
2112 {
2113         struct btrfs_trans_handle *trans;
2114         struct btrfs_fs_info *info = root->fs_info;
2115         struct btrfs_root *chunk_root = info->chunk_root;
2116         struct extent_buffer *leaf;
2117         struct btrfs_key key;
2118         struct btrfs_path path;
2119         struct btrfs_chunk chunk;
2120         unsigned long ptr;
2121         u32 size;
2122         u64 type;
2123         int ret;
2124
2125         btrfs_init_path(&path);
2126
2127         trans = btrfs_start_transaction(root, 1);
2128         BUG_ON(!trans);
2129
2130         /*
2131          * recow the whole chunk tree. this will move all chunk tree blocks
2132          * into system block group.
2133          */
2134         memset(&key, 0, sizeof(key));
2135         while (1) {
2136                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2137                 if (ret < 0)
2138                         goto err;
2139
2140                 ret = btrfs_next_leaf(chunk_root, &path);
2141                 if (ret < 0)
2142                         goto err;
2143                 if (ret > 0)
2144                         break;
2145
2146                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2147                 btrfs_release_path(&path);
2148         }
2149         btrfs_release_path(&path);
2150
2151         /* fixup the system chunk array in super block */
2152         btrfs_set_super_sys_array_size(info->super_copy, 0);
2153
2154         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2155         key.offset = 0;
2156         key.type = BTRFS_CHUNK_ITEM_KEY;
2157
2158         ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 0);
2159         if (ret < 0)
2160                 goto err;
2161         BUG_ON(ret != 0);
2162         while(1) {
2163                 leaf = path.nodes[0];
2164                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2165                         ret = btrfs_next_leaf(chunk_root, &path);
2166                         if (ret < 0)
2167                                 goto err;
2168                         if (ret > 0)
2169                                 break;
2170                         leaf = path.nodes[0];
2171                 }
2172                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2173                 if (key.type != BTRFS_CHUNK_ITEM_KEY)
2174                         goto next;
2175
2176                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2177                 size = btrfs_item_size_nr(leaf, path.slots[0]);
2178                 BUG_ON(size != sizeof(chunk));
2179                 read_extent_buffer(leaf, &chunk, ptr, size);
2180                 type = btrfs_stack_chunk_type(&chunk);
2181
2182                 if (!(type & BTRFS_BLOCK_GROUP_SYSTEM))
2183                         goto next;
2184
2185                 ret = btrfs_add_system_chunk(trans, chunk_root, &key,
2186                                              &chunk, size);
2187                 if (ret)
2188                         goto err;
2189 next:
2190                 path.slots[0]++;
2191         }
2192
2193         ret = btrfs_commit_transaction(trans, root);
2194         BUG_ON(ret);
2195 err:
2196         btrfs_release_path(&path);
2197         return ret;
2198 }
2199
2200 static int do_convert(const char *devname, int datacsum, int packing,
2201                 int noxattr)
2202 {
2203         int i, ret;
2204         int fd = -1;
2205         u32 blocksize;
2206         u64 blocks[7];
2207         u64 total_bytes;
2208         u64 super_bytenr;
2209         ext2_filsys ext2_fs;
2210         struct btrfs_root *root;
2211         struct btrfs_root *ext2_root;
2212
2213         ret = open_ext2fs(devname, &ext2_fs);
2214         if (ret) {
2215                 fprintf(stderr, "unable to open the Ext2fs\n");
2216                 goto fail;
2217         }
2218         blocksize = ext2_fs->blocksize;
2219         total_bytes = (u64)ext2_fs->super->s_blocks_count * blocksize;
2220         if (blocksize < 4096) {
2221                 fprintf(stderr, "block size is too small\n");
2222                 goto fail;
2223         }
2224         if (!(ext2_fs->super->s_feature_incompat &
2225               EXT2_FEATURE_INCOMPAT_FILETYPE)) {
2226                 fprintf(stderr, "filetype feature is missing\n");
2227                 goto fail;
2228         }
2229         for (i = 0; i < 7; i++) {
2230                 ret = ext2_alloc_block(ext2_fs, 0, blocks + i);
2231                 if (ret) {
2232                         fprintf(stderr, "not enough free space\n");
2233                         goto fail;
2234                 }
2235                 blocks[i] *= blocksize;
2236         }
2237         super_bytenr = blocks[0];
2238         fd = open(devname, O_RDWR);
2239         if (fd < 0) {
2240                 fprintf(stderr, "unable to open %s\n", devname);
2241                 goto fail;
2242         }
2243         ret = make_btrfs(fd, devname, ext2_fs->super->s_volume_name,
2244                          blocks, total_bytes, blocksize, blocksize,
2245                          blocksize, blocksize, 0);
2246         if (ret) {
2247                 fprintf(stderr, "unable to create initial ctree: %s\n",
2248                         strerror(-ret));
2249                 goto fail;
2250         }
2251         /* create a system chunk that maps the whole device */
2252         ret = prepare_system_chunk(fd, super_bytenr, blocksize);
2253         if (ret) {
2254                 fprintf(stderr, "unable to update system chunk\n");
2255                 goto fail;
2256         }
2257         root = open_ctree_fd(fd, devname, super_bytenr, OPEN_CTREE_WRITES);
2258         if (!root) {
2259                 fprintf(stderr, "unable to open ctree\n");
2260                 goto fail;
2261         }
2262         ret = cache_free_extents(root, ext2_fs);
2263         if (ret) {
2264                 fprintf(stderr, "error during cache_free_extents %d\n", ret);
2265                 goto fail;
2266         }
2267         root->fs_info->extent_ops = &extent_ops;
2268         /* recover block allocation bitmap */
2269         for (i = 0; i < 7; i++) {
2270                 blocks[i] /= blocksize;
2271                 ext2_free_block(ext2_fs, blocks[i]);
2272         }
2273         ret = init_btrfs(root);
2274         if (ret) {
2275                 fprintf(stderr, "unable to setup the root tree\n");
2276                 goto fail;
2277         }
2278         printf("creating btrfs metadata.\n");
2279         ret = copy_inodes(root, ext2_fs, datacsum, packing, noxattr);
2280         if (ret) {
2281                 fprintf(stderr, "error during copy_inodes %d\n", ret);
2282                 goto fail;
2283         }
2284         printf("creating ext2fs image file.\n");
2285         ext2_root = link_subvol(root, "ext2_saved", EXT2_IMAGE_SUBVOL_OBJECTID);
2286         if (!ext2_root) {
2287                 fprintf(stderr, "unable to create subvol\n");
2288                 goto fail;
2289         }
2290         ret = create_ext2_image(ext2_root, ext2_fs, "image");
2291         if (ret) {
2292                 fprintf(stderr, "error during create_ext2_image %d\n", ret);
2293                 goto fail;
2294         }
2295         printf("cleaning up system chunk.\n");
2296         ret = cleanup_sys_chunk(root, ext2_root);
2297         if (ret) {
2298                 fprintf(stderr, "error during cleanup_sys_chunk %d\n", ret);
2299                 goto fail;
2300         }
2301         ret = close_ctree(root);
2302         if (ret) {
2303                 fprintf(stderr, "error during close_ctree %d\n", ret);
2304                 goto fail;
2305         }
2306         close_ext2fs(ext2_fs);
2307
2308         /*
2309          * If this step succeed, we get a mountable btrfs. Otherwise
2310          * the ext2fs is left unchanged.
2311          */
2312         ret = migrate_super_block(fd, super_bytenr, blocksize);
2313         if (ret) {
2314                 fprintf(stderr, "unable to migrate super block\n");
2315                 goto fail;
2316         }
2317
2318         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2319         if (!root) {
2320                 fprintf(stderr, "unable to open ctree\n");
2321                 goto fail;
2322         }
2323         /* move chunk tree into system chunk. */
2324         ret = fixup_chunk_mapping(root);
2325         if (ret) {
2326                 fprintf(stderr, "error during fixup_chunk_tree\n");
2327                 goto fail;
2328         }
2329         ret = close_ctree(root);
2330         close(fd);
2331
2332         printf("conversion complete.\n");
2333         return 0;
2334 fail:
2335         if (fd != -1)
2336                 close(fd);
2337         fprintf(stderr, "conversion aborted.\n");
2338         return -1;
2339 }
2340
2341 static int may_rollback(struct btrfs_root *root)
2342 {
2343         struct btrfs_fs_info *info = root->fs_info;
2344         struct btrfs_multi_bio *multi = NULL;
2345         u64 bytenr;
2346         u64 length;
2347         u64 physical;
2348         u64 total_bytes;
2349         int num_stripes;
2350         int ret;
2351
2352         if (btrfs_super_num_devices(info->super_copy) != 1)
2353                 goto fail;
2354
2355         bytenr = BTRFS_SUPER_INFO_OFFSET;
2356         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
2357
2358         while (1) {
2359                 ret = btrfs_map_block(&info->mapping_tree, WRITE, bytenr,
2360                                       &length, &multi, 0, NULL);
2361                 if (ret)
2362                         goto fail;
2363
2364                 num_stripes = multi->num_stripes;
2365                 physical = multi->stripes[0].physical;
2366                 kfree(multi);
2367
2368                 if (num_stripes != 1 || physical != bytenr)
2369                         goto fail;
2370
2371                 bytenr += length;
2372                 if (bytenr >= total_bytes)
2373                         break;
2374         }
2375         return 0;
2376 fail:
2377         return -1;
2378 }
2379
2380 static int do_rollback(const char *devname, int force)
2381 {
2382         int fd = -1;
2383         int ret;
2384         int i;
2385         struct btrfs_root *root;
2386         struct btrfs_root *ext2_root;
2387         struct btrfs_root *chunk_root;
2388         struct btrfs_dir_item *dir;
2389         struct btrfs_inode_item *inode;
2390         struct btrfs_file_extent_item *fi;
2391         struct btrfs_trans_handle *trans;
2392         struct extent_buffer *leaf;
2393         struct btrfs_block_group_cache *cache1;
2394         struct btrfs_block_group_cache *cache2;
2395         struct btrfs_key key;
2396         struct btrfs_path path;
2397         struct extent_io_tree io_tree;
2398         char *buf = NULL;
2399         char *name;
2400         u64 bytenr;
2401         u64 num_bytes;
2402         u64 root_dir;
2403         u64 objectid;
2404         u64 offset;
2405         u64 start;
2406         u64 end;
2407         u64 sb_bytenr;
2408         u64 first_free;
2409         u64 total_bytes;
2410         u32 sectorsize;
2411
2412         extent_io_tree_init(&io_tree);
2413
2414         fd = open(devname, O_RDWR);
2415         if (fd < 0) {
2416                 fprintf(stderr, "unable to open %s\n", devname);
2417                 goto fail;
2418         }
2419         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2420         if (!root) {
2421                 fprintf(stderr, "unable to open ctree\n");
2422                 goto fail;
2423         }
2424         ret = may_rollback(root);
2425         if (ret < 0) {
2426                 fprintf(stderr, "unable to do rollback\n");
2427                 goto fail;
2428         }
2429
2430         sectorsize = root->sectorsize;
2431         buf = malloc(sectorsize);
2432         if (!buf) {
2433                 fprintf(stderr, "unable to allocate memory\n");
2434                 goto fail;
2435         }
2436
2437         btrfs_init_path(&path);
2438
2439         key.objectid = EXT2_IMAGE_SUBVOL_OBJECTID;
2440         key.type = BTRFS_ROOT_ITEM_KEY;
2441         key.offset = (u64)-1;
2442         ext2_root = btrfs_read_fs_root(root->fs_info, &key);
2443         if (!ext2_root || IS_ERR(ext2_root)) {
2444                 fprintf(stderr, "unable to open subvol %llu\n",
2445                         key.objectid);
2446                 goto fail;
2447         }
2448
2449         name = "image";
2450         root_dir = btrfs_root_dirid(&root->root_item);
2451         dir = btrfs_lookup_dir_item(NULL, ext2_root, &path,
2452                                    root_dir, name, strlen(name), 0);
2453         if (!dir || IS_ERR(dir)) {
2454                 fprintf(stderr, "unable to find file %s\n", name);
2455                 goto fail;
2456         }
2457         leaf = path.nodes[0];
2458         btrfs_dir_item_key_to_cpu(leaf, dir, &key);
2459         btrfs_release_path(&path);
2460
2461         objectid = key.objectid;
2462
2463         ret = btrfs_lookup_inode(NULL, ext2_root, &path, &key, 0);
2464         if (ret) {
2465                 fprintf(stderr, "unable to find inode item\n");
2466                 goto fail;
2467         }
2468         leaf = path.nodes[0];
2469         inode = btrfs_item_ptr(leaf, path.slots[0], struct btrfs_inode_item);
2470         total_bytes = btrfs_inode_size(leaf, inode);
2471         btrfs_release_path(&path);
2472
2473         key.objectid = objectid;
2474         key.offset = 0;
2475         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
2476         ret = btrfs_search_slot(NULL, ext2_root, &key, &path, 0, 0);
2477         if (ret != 0) {
2478                 fprintf(stderr, "unable to find first file extent\n");
2479                 btrfs_release_path(&path);
2480                 goto fail;
2481         }
2482
2483         /* build mapping tree for the relocated blocks */
2484         for (offset = 0; offset < total_bytes; ) {
2485                 leaf = path.nodes[0];
2486                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2487                         ret = btrfs_next_leaf(root, &path);
2488                         if (ret != 0)
2489                                 break;
2490                         continue;
2491                 }
2492
2493                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2494                 if (key.objectid != objectid || key.offset != offset ||
2495                     btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2496                         break;
2497
2498                 fi = btrfs_item_ptr(leaf, path.slots[0],
2499                                     struct btrfs_file_extent_item);
2500                 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2501                         break;
2502                 if (btrfs_file_extent_compression(leaf, fi) ||
2503                     btrfs_file_extent_encryption(leaf, fi) ||
2504                     btrfs_file_extent_other_encoding(leaf, fi))
2505                         break;
2506
2507                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2508                 /* skip holes and direct mapped extents */
2509                 if (bytenr == 0 || bytenr == offset)
2510                         goto next_extent;
2511
2512                 bytenr += btrfs_file_extent_offset(leaf, fi);
2513                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
2514
2515                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2516                 cache2 =  btrfs_lookup_block_group(root->fs_info,
2517                                                    offset + num_bytes - 1);
2518                 if (!cache1 || cache1 != cache2 ||
2519                     (!(cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
2520                      !intersect_with_sb(offset, num_bytes)))
2521                         break;
2522
2523                 set_extent_bits(&io_tree, offset, offset + num_bytes - 1,
2524                                 EXTENT_LOCKED, GFP_NOFS);
2525                 set_state_private(&io_tree, offset, bytenr);
2526 next_extent:
2527                 offset += btrfs_file_extent_num_bytes(leaf, fi);
2528                 path.slots[0]++;
2529         }
2530         btrfs_release_path(&path);
2531
2532         if (offset < total_bytes) {
2533                 fprintf(stderr, "unable to build extent mapping\n");
2534                 goto fail;
2535         }
2536
2537         first_free = BTRFS_SUPER_INFO_OFFSET + 2 * sectorsize - 1;
2538         first_free &= ~((u64)sectorsize - 1);
2539         /* backup for extent #0 should exist */
2540         if(!test_range_bit(&io_tree, 0, first_free - 1, EXTENT_LOCKED, 1)) {
2541                 fprintf(stderr, "no backup for the first extent\n");
2542                 goto fail;
2543         }
2544         /* force no allocation from system block group */
2545         root->fs_info->system_allocs = -1;
2546         trans = btrfs_start_transaction(root, 1);
2547         BUG_ON(!trans);
2548         /*
2549          * recow the whole chunk tree, this will remove all chunk tree blocks
2550          * from system block group
2551          */
2552         chunk_root = root->fs_info->chunk_root;
2553         memset(&key, 0, sizeof(key));
2554         while (1) {
2555                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2556                 if (ret < 0)
2557                         break;
2558
2559                 ret = btrfs_next_leaf(chunk_root, &path);
2560                 if (ret)
2561                         break;
2562
2563                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2564                 btrfs_release_path(&path);
2565         }
2566         btrfs_release_path(&path);
2567
2568         offset = 0;
2569         num_bytes = 0;
2570         while(1) {
2571                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2572                 if (!cache1)
2573                         break;
2574
2575                 if (cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2576                         num_bytes += btrfs_block_group_used(&cache1->item);
2577
2578                 offset = cache1->key.objectid + cache1->key.offset;
2579         }
2580         /* only extent #0 left in system block group? */
2581         if (num_bytes > first_free) {
2582                 fprintf(stderr, "unable to empty system block group\n");
2583                 goto fail;
2584         }
2585         /* create a system chunk that maps the whole device */
2586         ret = prepare_system_chunk_sb(root->fs_info->super_copy);
2587         if (ret) {
2588                 fprintf(stderr, "unable to update system chunk\n");
2589                 goto fail;
2590         }
2591
2592         ret = btrfs_commit_transaction(trans, root);
2593         BUG_ON(ret);
2594
2595         ret = close_ctree(root);
2596         if (ret) {
2597                 fprintf(stderr, "error during close_ctree %d\n", ret);
2598                 goto fail;
2599         }
2600
2601         /* zero btrfs super block mirrors */
2602         memset(buf, 0, sectorsize);
2603         for (i = 1 ; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2604                 bytenr = btrfs_sb_offset(i);
2605                 if (bytenr >= total_bytes)
2606                         break;
2607                 ret = pwrite(fd, buf, sectorsize, bytenr);
2608         }
2609
2610         sb_bytenr = (u64)-1;
2611         /* copy all relocated blocks back */
2612         while(1) {
2613                 ret = find_first_extent_bit(&io_tree, 0, &start, &end,
2614                                             EXTENT_LOCKED);
2615                 if (ret)
2616                         break;
2617
2618                 ret = get_state_private(&io_tree, start, &bytenr);
2619                 BUG_ON(ret);
2620
2621                 clear_extent_bits(&io_tree, start, end, EXTENT_LOCKED,
2622                                   GFP_NOFS);
2623
2624                 while (start <= end) {
2625                         if (start == BTRFS_SUPER_INFO_OFFSET) {
2626                                 sb_bytenr = bytenr;
2627                                 goto next_sector;
2628                         }
2629                         ret = pread(fd, buf, sectorsize, bytenr);
2630                         if (ret < 0) {
2631                                 fprintf(stderr, "error during pread %d\n", ret);
2632                                 goto fail;
2633                         }
2634                         BUG_ON(ret != sectorsize);
2635                         ret = pwrite(fd, buf, sectorsize, start);
2636                         if (ret < 0) {
2637                                 fprintf(stderr, "error during pwrite %d\n", ret);
2638                                 goto fail;
2639                         }
2640                         BUG_ON(ret != sectorsize);
2641 next_sector:
2642                         start += sectorsize;
2643                         bytenr += sectorsize;
2644                 }
2645         }
2646
2647         ret = fsync(fd);
2648         if (ret) {
2649                 fprintf(stderr, "error during fsync %d\n", ret);
2650                 goto fail;
2651         }
2652         /*
2653          * finally, overwrite btrfs super block.
2654          */
2655         ret = pread(fd, buf, sectorsize, sb_bytenr);
2656         if (ret < 0) {
2657                 fprintf(stderr, "error during pread %d\n", ret);
2658                 goto fail;
2659         }
2660         BUG_ON(ret != sectorsize);
2661         ret = pwrite(fd, buf, sectorsize, BTRFS_SUPER_INFO_OFFSET);
2662         if (ret < 0) {
2663                 fprintf(stderr, "error during pwrite %d\n", ret);
2664                 goto fail;
2665         }
2666         BUG_ON(ret != sectorsize);
2667         ret = fsync(fd);
2668         if (ret) {
2669                 fprintf(stderr, "error during fsync %d\n", ret);
2670                 goto fail;
2671         }
2672
2673         close(fd);
2674         free(buf);
2675         extent_io_tree_cleanup(&io_tree);
2676         printf("rollback complete.\n");
2677         return 0;
2678
2679 fail:
2680         if (fd != -1)
2681                 close(fd);
2682         free(buf);
2683         fprintf(stderr, "rollback aborted.\n");
2684         return -1;
2685 }
2686
2687 static void print_usage(void)
2688 {
2689         printf("usage: btrfs-convert [-d] [-i] [-n] [-r] device\n");
2690         printf("\t-d disable data checksum\n");
2691         printf("\t-i ignore xattrs and ACLs\n");
2692         printf("\t-n disable packing of small files\n");
2693         printf("\t-r roll back to ext2fs\n");
2694 }
2695
2696 int main(int argc, char *argv[])
2697 {
2698         int ret;
2699         int packing = 1;
2700         int noxattr = 0;
2701         int datacsum = 1;
2702         int rollback = 0;
2703         char *file;
2704         while(1) {
2705                 int c = getopt(argc, argv, "dinr");
2706                 if (c < 0)
2707                         break;
2708                 switch(c) {
2709                         case 'd':
2710                                 datacsum = 0;
2711                                 break;
2712                         case 'i':
2713                                 noxattr = 1;
2714                                 break;
2715                         case 'n':
2716                                 packing = 0;
2717                                 break;
2718                         case 'r':
2719                                 rollback = 1;
2720                                 break;
2721                         default:
2722                                 print_usage();
2723                                 return 1;
2724                 }
2725         }
2726         argc = argc - optind;
2727         if (argc != 1) {
2728                 print_usage();
2729                 return 1;
2730         }
2731
2732         file = argv[optind];
2733         if (check_mounted(file)) {
2734                 fprintf(stderr, "%s is mounted\n", file);
2735                 return 1;
2736         }
2737
2738         if (rollback) {
2739                 ret = do_rollback(file, 0);
2740         } else {
2741                 ret = do_convert(file, datacsum, packing, noxattr);
2742         }
2743         if (ret)
2744                 return 1;
2745         return 0;
2746 }