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 "task-utils.h"
  42 #include <ext2fs/ext2_fs.h>
  43 #include <ext2fs/ext2fs.h>
  44 #include <ext2fs/ext2_ext_attr.h>
  45
  46 #define INO_OFFSET (BTRFS_FIRST_FREE_OBJECTID - EXT2_ROOT_INO)
  47 #define EXT2_IMAGE_SUBVOL_OBJECTID BTRFS_FIRST_FREE_OBJECTID
  48
  49 struct task_ctx {
  50         uint32_t max_copy_inodes;
  51         uint32_t cur_copy_inodes;
  52         struct task_info *info;
  53 };
  54
  55 static void *print_copied_inodes(void *p)
  56 {
  57         struct task_ctx *priv = p;
  58         const char work_indicator[] = { '.', 'o', 'O', 'o' };
  59         uint32_t count = 0;
  60
  61         task_period_start(priv->info, 1000 /* 1s */);
  62         while (1) {
  63                 count++;
  64                 printf("copy inodes [%c] [%10d/%10d]\r",
  65                        work_indicator[count % 4], priv->cur_copy_inodes,
  66                        priv->max_copy_inodes);
  67                 fflush(stdout);
  68                 task_period_wait(priv->info);
  69         }
  70
  71         return NULL;
  72 }
  73
  74 static int after_copied_inodes(void *p)
  75 {
  76         struct task_ctx *priv = p;
  77
  78         printf("\n");
  79         task_period_stop(priv->info);
  80
  81         return 0;
  82 }
  83
  84 /*
  85  * Open Ext2fs in readonly mode, read block allocation bitmap and
  86  * inode bitmap into memory.
  87  */
  88 static int open_ext2fs(const char *name, ext2_filsys *ret_fs)
  89 {
  90         errcode_t ret;
  91         ext2_filsys ext2_fs;
  92         ext2_ino_t ino;
  93         ret = ext2fs_open(name, 0, 0, 0, unix_io_manager, &ext2_fs);
  94         if (ret) {
  95                 fprintf(stderr, "ext2fs_open: %s\n", error_message(ret));
  96                 goto fail;
  97         }
  98         ret = ext2fs_read_inode_bitmap(ext2_fs);
  99         if (ret) {
 100                 fprintf(stderr, "ext2fs_read_inode_bitmap: %s\n",
 101                         error_message(ret));
 102                 goto fail;
 103         }
 104         ret = ext2fs_read_block_bitmap(ext2_fs);
 105         if (ret) {
 106                 fprintf(stderr, "ext2fs_read_block_bitmap: %s\n",
 107                         error_message(ret));
 108                 goto fail;
 109         }
 110         /*
 111          * search each block group for a free inode. this set up
 112          * uninit block/inode bitmaps appropriately.
 113          */
 114         ino = 1;
 115         while (ino <= ext2_fs->super->s_inodes_count) {
 116                 ext2_ino_t foo;
 117                 ext2fs_new_inode(ext2_fs, ino, 0, NULL, &foo);
 118                 ino += EXT2_INODES_PER_GROUP(ext2_fs->super);
 119         }
 120
 121         *ret_fs = ext2_fs;
 122         return 0;
 123 fail:
 124         return -1;
 125 }
 126
 127 static int close_ext2fs(ext2_filsys fs)
 128 {
 129         ext2fs_close(fs);
 130         return 0;
 131 }
 132
 133 static int ext2_alloc_block(ext2_filsys fs, u64 goal, u64 *block_ret)
 134 {
 135         blk_t block;
 136
 137         if (!ext2fs_new_block(fs, goal, NULL, &block)) {
 138                 ext2fs_fast_mark_block_bitmap(fs->block_map, block);
 139                 *block_ret = block;
 140                 return 0;
 141         }
 142         return -ENOSPC;
 143 }
 144
 145 static int ext2_free_block(ext2_filsys fs, u64 block)
 146 {
 147         BUG_ON(block != (blk_t)block);
 148         ext2fs_fast_unmark_block_bitmap(fs->block_map, block);
 149         return 0;
 150 }
 151
 152 static int cache_free_extents(struct btrfs_root *root, ext2_filsys ext2_fs)
 153
 154 {
 155         int i, ret = 0;
 156         blk_t block;
 157         u64 bytenr;
 158         u64 blocksize = ext2_fs->blocksize;
 159
 160         block = ext2_fs->super->s_first_data_block;
 161         for (; block < ext2_fs->super->s_blocks_count; block++) {
 162                 if (ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
 163                         continue;
 164                 bytenr = block * blocksize;
 165                 ret = set_extent_dirty(&root->fs_info->free_space_cache,
 166                                        bytenr, bytenr + blocksize - 1, 0);
 167                 BUG_ON(ret);
 168         }
 169
 170         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 171                 bytenr = btrfs_sb_offset(i);
 172                 bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 173                 if (bytenr >= blocksize * ext2_fs->super->s_blocks_count)
 174                         break;
 175                 clear_extent_dirty(&root->fs_info->free_space_cache, bytenr,
 176                                    bytenr + BTRFS_STRIPE_LEN - 1, 0);
 177         }
 178
 179         clear_extent_dirty(&root->fs_info->free_space_cache,
 180                            0, BTRFS_SUPER_INFO_OFFSET - 1, 0);
 181
 182         return 0;
 183 }
 184
 185 static int custom_alloc_extent(struct btrfs_root *root, u64 num_bytes,
 186                                u64 hint_byte, struct btrfs_key *ins)
 187 {
 188         u64 start;
 189         u64 end;
 190         u64 last = hint_byte;
 191         int ret;
 192         int wrapped = 0;
 193         struct btrfs_block_group_cache *cache;
 194
 195         while(1) {
 196                 ret = find_first_extent_bit(&root->fs_info->free_space_cache,
 197                                             last, &start, &end, EXTENT_DIRTY);
 198                 if (ret) {
 199                         if (wrapped++ == 0) {
 200                                 last = 0;
 201                                 continue;
 202                         } else {
 203                                 goto fail;
 204                         }
 205                 }
 206
 207                 start = max(last, start);
 208                 last = end + 1;
 209                 if (last - start < num_bytes)
 210                         continue;
 211
 212                 last = start + num_bytes;
 213                 if (test_range_bit(&root->fs_info->pinned_extents,
 214                                    start, last - 1, EXTENT_DIRTY, 0))
 215                         continue;
 216
 217                 cache = btrfs_lookup_block_group(root->fs_info, start);
 218                 BUG_ON(!cache);
 219                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM ||
 220                     last > cache->key.objectid + cache->key.offset) {
 221                         last = cache->key.objectid + cache->key.offset;
 222                         continue;
 223                 }
 224
 225                 clear_extent_dirty(&root->fs_info->free_space_cache,
 226                                    start, start + num_bytes - 1, 0);
 227
 228                 ins->objectid = start;
 229                 ins->offset = num_bytes;
 230                 ins->type = BTRFS_EXTENT_ITEM_KEY;
 231                 return 0;
 232         }
 233 fail:
 234         fprintf(stderr, "not enough free space\n");
 235         return -ENOSPC;
 236 }
 237
 238 static int intersect_with_sb(u64 bytenr, u64 num_bytes)
 239 {
 240         int i;
 241         u64 offset;
 242
 243         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 244                 offset = btrfs_sb_offset(i);
 245                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
 246
 247                 if (bytenr < offset + BTRFS_STRIPE_LEN &&
 248                     bytenr + num_bytes > offset)
 249                         return 1;
 250         }
 251         return 0;
 252 }
 253
 254 static int custom_free_extent(struct btrfs_root *root, u64 bytenr,
 255                               u64 num_bytes)
 256 {
 257         return intersect_with_sb(bytenr, num_bytes);
 258 }
 259
 260 static struct btrfs_extent_ops extent_ops = {
 261         .alloc_extent = custom_alloc_extent,
 262         .free_extent = custom_free_extent,
 263 };
 264
 265 struct dir_iterate_data {
 266         struct btrfs_trans_handle *trans;
 267         struct btrfs_root *root;
 268         struct btrfs_inode_item *inode;
 269         u64 objectid;
 270         u64 index_cnt;
 271         u64 parent;
 272         int errcode;
 273 };
 274
 275 static u8 filetype_conversion_table[EXT2_FT_MAX] = {
 276         [EXT2_FT_UNKNOWN]       = BTRFS_FT_UNKNOWN,
 277         [EXT2_FT_REG_FILE]      = BTRFS_FT_REG_FILE,
 278         [EXT2_FT_DIR]           = BTRFS_FT_DIR,
 279         [EXT2_FT_CHRDEV]        = BTRFS_FT_CHRDEV,
 280         [EXT2_FT_BLKDEV]        = BTRFS_FT_BLKDEV,
 281         [EXT2_FT_FIFO]          = BTRFS_FT_FIFO,
 282         [EXT2_FT_SOCK]          = BTRFS_FT_SOCK,
 283         [EXT2_FT_SYMLINK]       = BTRFS_FT_SYMLINK,
 284 };
 285
 286 static int dir_iterate_proc(ext2_ino_t dir, int entry,
 287                             struct ext2_dir_entry *dirent,
 288                             int offset, int blocksize,
 289                             char *buf,void *priv_data)
 290 {
 291         int ret;
 292         int file_type;
 293         u64 objectid;
 294         u64 inode_size;
 295         char dotdot[] = "..";
 296         struct btrfs_key location;
 297         struct dir_iterate_data *idata = (struct dir_iterate_data *)priv_data;
 298         int name_len;
 299
 300         name_len = dirent->name_len & 0xFF;
 301
 302         objectid = dirent->inode + INO_OFFSET;
 303         if (!strncmp(dirent->name, dotdot, name_len)) {
 304                 if (name_len == 2) {
 305                         BUG_ON(idata->parent != 0);
 306                         idata->parent = objectid;
 307                 }
 308                 return 0;
 309         }
 310         if (dirent->inode < EXT2_GOOD_OLD_FIRST_INO)
 311                 return 0;
 312
 313         location.objectid = objectid;
 314         location.offset = 0;
 315         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
 316
 317         file_type = dirent->name_len >> 8;
 318         BUG_ON(file_type > EXT2_FT_SYMLINK);
 319         ret = btrfs_insert_dir_item(idata->trans, idata->root,
 320                                     dirent->name, name_len,
 321                                     idata->objectid, &location,
 322                                     filetype_conversion_table[file_type],
 323                                     idata->index_cnt);
 324         if (ret)
 325                 goto fail;
 326         ret = btrfs_insert_inode_ref(idata->trans, idata->root,
 327                                      dirent->name, name_len,
 328                                      objectid, idata->objectid,
 329                                      idata->index_cnt);
 330         if (ret)
 331                 goto fail;
 332         idata->index_cnt++;
 333         inode_size = btrfs_stack_inode_size(idata->inode) +
 334                      name_len * 2;
 335         btrfs_set_stack_inode_size(idata->inode, inode_size);
 336         return 0;
 337 fail:
 338         idata->errcode = ret;
 339         return BLOCK_ABORT;
 340 }
 341
 342 static int create_dir_entries(struct btrfs_trans_handle *trans,
 343                               struct btrfs_root *root, u64 objectid,
 344                               struct btrfs_inode_item *btrfs_inode,
 345                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 346 {
 347         int ret;
 348         errcode_t err;
 349         struct dir_iterate_data data = {
 350                 .trans          = trans,
 351                 .root           = root,
 352                 .inode          = btrfs_inode,
 353                 .objectid       = objectid,
 354                 .index_cnt      = 2,
 355                 .parent         = 0,
 356                 .errcode        = 0,
 357         };
 358
 359         err = ext2fs_dir_iterate2(ext2_fs, ext2_ino, 0, NULL,
 360                                   dir_iterate_proc, &data);
 361         if (err)
 362                 goto error;
 363         ret = data.errcode;
 364         if (ret == 0 && data.parent == objectid) {
 365                 ret = btrfs_insert_inode_ref(trans, root, "..", 2,
 366                                              objectid, objectid, 0);
 367         }
 368         return ret;
 369 error:
 370         fprintf(stderr, "ext2fs_dir_iterate2: %s\n", error_message(err));
 371         return -1;
 372 }
 373
 374 static int read_disk_extent(struct btrfs_root *root, u64 bytenr,
 375                             u32 num_bytes, char *buffer)
 376 {
 377         int ret;
 378         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
 379
 380         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, bytenr);
 381         if (ret != num_bytes)
 382                 goto fail;
 383         ret = 0;
 384 fail:
 385         if (ret > 0)
 386                 ret = -1;
 387         return ret;
 388 }
 389
 390 static int csum_disk_extent(struct btrfs_trans_handle *trans,
 391                             struct btrfs_root *root,
 392                             u64 disk_bytenr, u64 num_bytes)
 393 {
 394         u32 blocksize = root->sectorsize;
 395         u64 offset;
 396         char *buffer;
 397         int ret = 0;
 398
 399         buffer = malloc(blocksize);
 400         if (!buffer)
 401                 return -ENOMEM;
 402         for (offset = 0; offset < num_bytes; offset += blocksize) {
 403                 ret = read_disk_extent(root, disk_bytenr + offset,
 404                                         blocksize, buffer);
 405                 if (ret)
 406                         break;
 407                 ret = btrfs_csum_file_block(trans,
 408                                             root->fs_info->csum_root,
 409                                             disk_bytenr + num_bytes,
 410                                             disk_bytenr + offset,
 411                                             buffer, blocksize);
 412                 if (ret)
 413                         break;
 414         }
 415         free(buffer);
 416         return ret;
 417 }
 418
 419 static int record_file_blocks(struct btrfs_trans_handle *trans,
 420                               struct btrfs_root *root, u64 objectid,
 421                               struct btrfs_inode_item *inode,
 422                               u64 file_block, u64 disk_block,
 423                               u64 num_blocks, int checksum)
 424 {
 425         int ret;
 426         u64 file_pos = file_block * root->sectorsize;
 427         u64 disk_bytenr = disk_block * root->sectorsize;
 428         u64 num_bytes = num_blocks * root->sectorsize;
 429         ret = btrfs_record_file_extent(trans, root, objectid, inode, file_pos,
 430                                         disk_bytenr, num_bytes);
 431
 432         if (ret || !checksum || disk_bytenr == 0)
 433                 return ret;
 434
 435         return csum_disk_extent(trans, root, disk_bytenr, num_bytes);
 436 }
 437
 438 struct blk_iterate_data {
 439         struct btrfs_trans_handle *trans;
 440         struct btrfs_root *root;
 441         struct btrfs_inode_item *inode;
 442         u64 objectid;
 443         u64 first_block;
 444         u64 disk_block;
 445         u64 num_blocks;
 446         u64 boundary;
 447         int checksum;
 448         int errcode;
 449 };
 450
 451 static int block_iterate_proc(ext2_filsys ext2_fs,
 452                               u64 disk_block, u64 file_block,
 453                               struct blk_iterate_data *idata)
 454 {
 455         int ret;
 456         int sb_region;
 457         int do_barrier;
 458         struct btrfs_root *root = idata->root;
 459         struct btrfs_trans_handle *trans = idata->trans;
 460         struct btrfs_block_group_cache *cache;
 461         u64 bytenr = disk_block * root->sectorsize;
 462
 463         sb_region = intersect_with_sb(bytenr, root->sectorsize);
 464         do_barrier = sb_region || disk_block >= idata->boundary;
 465         if ((idata->num_blocks > 0 && do_barrier) ||
 466             (file_block > idata->first_block + idata->num_blocks) ||
 467             (disk_block != idata->disk_block + idata->num_blocks)) {
 468                 if (idata->num_blocks > 0) {
 469                         ret = record_file_blocks(trans, root, idata->objectid,
 470                                         idata->inode, idata->first_block,
 471                                         idata->disk_block, idata->num_blocks,
 472                                         idata->checksum);
 473                         if (ret)
 474                                 goto fail;
 475                         idata->first_block += idata->num_blocks;
 476                         idata->num_blocks = 0;
 477                 }
 478                 if (file_block > idata->first_block) {
 479                         ret = record_file_blocks(trans, root, idata->objectid,
 480                                         idata->inode, idata->first_block,
 481                                         0, file_block - idata->first_block,
 482                                         idata->checksum);
 483                         if (ret)
 484                                 goto fail;
 485                 }
 486
 487                 if (sb_region) {
 488                         bytenr += BTRFS_STRIPE_LEN - 1;
 489                         bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 490                 } else {
 491                         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
 492                         BUG_ON(!cache);
 493                         bytenr = cache->key.objectid + cache->key.offset;
 494                 }
 495
 496                 idata->first_block = file_block;
 497                 idata->disk_block = disk_block;
 498                 idata->boundary = bytenr / root->sectorsize;
 499         }
 500         idata->num_blocks++;
 501         return 0;
 502 fail:
 503         idata->errcode = ret;
 504         return BLOCK_ABORT;
 505 }
 506
 507 static int __block_iterate_proc(ext2_filsys fs, blk_t *blocknr,
 508                                 e2_blkcnt_t blockcnt, blk_t ref_block,
 509                                 int ref_offset, void *priv_data)
 510 {
 511         struct blk_iterate_data *idata;
 512         idata = (struct blk_iterate_data *)priv_data;
 513         return block_iterate_proc(fs, *blocknr, blockcnt, idata);
 514 }
 515
 516 /*
 517  * traverse file's data blocks, record these data blocks as file extents.
 518  */
 519 static int create_file_extents(struct btrfs_trans_handle *trans,
 520                                struct btrfs_root *root, u64 objectid,
 521                                struct btrfs_inode_item *btrfs_inode,
 522                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 523                                int datacsum, int packing)
 524 {
 525         int ret;
 526         char *buffer = NULL;
 527         errcode_t err;
 528         u32 last_block;
 529         u32 sectorsize = root->sectorsize;
 530         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 531         struct blk_iterate_data data = {
 532                 .trans          = trans,
 533                 .root           = root,
 534                 .inode          = btrfs_inode,
 535                 .objectid       = objectid,
 536                 .first_block    = 0,
 537                 .disk_block     = 0,
 538                 .num_blocks     = 0,
 539                 .boundary       = (u64)-1,
 540                 .checksum       = datacsum,
 541                 .errcode        = 0,
 542         };
 543         err = ext2fs_block_iterate2(ext2_fs, ext2_ino, BLOCK_FLAG_DATA_ONLY,
 544                                     NULL, __block_iterate_proc, &data);
 545         if (err)
 546                 goto error;
 547         ret = data.errcode;
 548         if (ret)
 549                 goto fail;
 550         if (packing && data.first_block == 0 && data.num_blocks > 0 &&
 551             inode_size <= BTRFS_MAX_INLINE_DATA_SIZE(root)) {
 552                 u64 num_bytes = data.num_blocks * sectorsize;
 553                 u64 disk_bytenr = data.disk_block * sectorsize;
 554                 u64 nbytes;
 555
 556                 buffer = malloc(num_bytes);
 557                 if (!buffer)
 558                         return -ENOMEM;
 559                 ret = read_disk_extent(root, disk_bytenr, num_bytes, buffer);
 560                 if (ret)
 561                         goto fail;
 562                 if (num_bytes > inode_size)
 563                         num_bytes = inode_size;
 564                 ret = btrfs_insert_inline_extent(trans, root, objectid,
 565                                                  0, buffer, num_bytes);
 566                 if (ret)
 567                         goto fail;
 568                 nbytes = btrfs_stack_inode_nbytes(btrfs_inode) + num_bytes;
 569                 btrfs_set_stack_inode_nbytes(btrfs_inode, nbytes);
 570         } else if (data.num_blocks > 0) {
 571                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 572                                          data.first_block, data.disk_block,
 573                                          data.num_blocks, data.checksum);
 574                 if (ret)
 575                         goto fail;
 576         }
 577         data.first_block += data.num_blocks;
 578         last_block = (inode_size + sectorsize - 1) / sectorsize;
 579         if (last_block > data.first_block) {
 580                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 581                                          data.first_block, 0, last_block -
 582                                          data.first_block, data.checksum);
 583         }
 584 fail:
 585         free(buffer);
 586         return ret;
 587 error:
 588         fprintf(stderr, "ext2fs_block_iterate2: %s\n", error_message(err));
 589         return -1;
 590 }
 591
 592 static int create_symbol_link(struct btrfs_trans_handle *trans,
 593                               struct btrfs_root *root, u64 objectid,
 594                               struct btrfs_inode_item *btrfs_inode,
 595                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 596                               struct ext2_inode *ext2_inode)
 597 {
 598         int ret;
 599         char *pathname;
 600         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 601         if (ext2fs_inode_data_blocks(ext2_fs, ext2_inode)) {
 602                 btrfs_set_stack_inode_size(btrfs_inode, inode_size + 1);
 603                 ret = create_file_extents(trans, root, objectid, btrfs_inode,
 604                                           ext2_fs, ext2_ino, 1, 1);
 605                 btrfs_set_stack_inode_size(btrfs_inode, inode_size);
 606                 return ret;
 607         }
 608
 609         pathname = (char *)&(ext2_inode->i_block[0]);
 610         BUG_ON(pathname[inode_size] != 0);
 611         ret = btrfs_insert_inline_extent(trans, root, objectid, 0,
 612                                          pathname, inode_size + 1);
 613         btrfs_set_stack_inode_nbytes(btrfs_inode, inode_size + 1);
 614         return ret;
 615 }
 616
 617 /*
 618  * Following xattr/acl related codes are based on codes in
 619  * fs/ext3/xattr.c and fs/ext3/acl.c
 620  */
 621 #define EXT2_XATTR_BHDR(ptr) ((struct ext2_ext_attr_header *)(ptr))
 622 #define EXT2_XATTR_BFIRST(ptr) \
 623         ((struct ext2_ext_attr_entry *)(EXT2_XATTR_BHDR(ptr) + 1))
 624 #define EXT2_XATTR_IHDR(inode) \
 625         ((struct ext2_ext_attr_header *) ((void *)(inode) + \
 626                 EXT2_GOOD_OLD_INODE_SIZE + (inode)->i_extra_isize))
 627 #define EXT2_XATTR_IFIRST(inode) \
 628         ((struct ext2_ext_attr_entry *) ((void *)EXT2_XATTR_IHDR(inode) + \
 629                 sizeof(EXT2_XATTR_IHDR(inode)->h_magic)))
 630
 631 static int ext2_xattr_check_names(struct ext2_ext_attr_entry *entry,
 632                                   const void *end)
 633 {
 634         struct ext2_ext_attr_entry *next;
 635
 636         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 637                 next = EXT2_EXT_ATTR_NEXT(entry);
 638                 if ((void *)next >= end)
 639                         return -EIO;
 640                 entry = next;
 641         }
 642         return 0;
 643 }
 644
 645 static int ext2_xattr_check_block(const char *buf, size_t size)
 646 {
 647         int error;
 648         struct ext2_ext_attr_header *header = EXT2_XATTR_BHDR(buf);
 649
 650         if (header->h_magic != EXT2_EXT_ATTR_MAGIC ||
 651             header->h_blocks != 1)
 652                 return -EIO;
 653         error = ext2_xattr_check_names(EXT2_XATTR_BFIRST(buf), buf + size);
 654         return error;
 655 }
 656
 657 static int ext2_xattr_check_entry(struct ext2_ext_attr_entry *entry,
 658                                   size_t size)
 659 {
 660         size_t value_size = entry->e_value_size;
 661
 662         if (entry->e_value_block != 0 || value_size > size ||
 663             entry->e_value_offs + value_size > size)
 664                 return -EIO;
 665         return 0;
 666 }
 667
 668 #define EXT2_ACL_VERSION        0x0001
 669
 670 typedef struct {
 671         __le16          e_tag;
 672         __le16          e_perm;
 673         __le32          e_id;
 674 } ext2_acl_entry;
 675
 676 typedef struct {
 677         __le16          e_tag;
 678         __le16          e_perm;
 679 } ext2_acl_entry_short;
 680
 681 typedef struct {
 682         __le32          a_version;
 683 } ext2_acl_header;
 684
 685 static inline int ext2_acl_count(size_t size)
 686 {
 687         ssize_t s;
 688         size -= sizeof(ext2_acl_header);
 689         s = size - 4 * sizeof(ext2_acl_entry_short);
 690         if (s < 0) {
 691                 if (size % sizeof(ext2_acl_entry_short))
 692                         return -1;
 693                 return size / sizeof(ext2_acl_entry_short);
 694         } else {
 695                 if (s % sizeof(ext2_acl_entry))
 696                         return -1;
 697                 return s / sizeof(ext2_acl_entry) + 4;
 698         }
 699 }
 700
 701 #define ACL_EA_VERSION          0x0002
 702
 703 typedef struct {
 704         __le16          e_tag;
 705         __le16          e_perm;
 706         __le32          e_id;
 707 } acl_ea_entry;
 708
 709 typedef struct {
 710         __le32          a_version;
 711         acl_ea_entry    a_entries[0];
 712 } acl_ea_header;
 713
 714 static inline size_t acl_ea_size(int count)
 715 {
 716         return sizeof(acl_ea_header) + count * sizeof(acl_ea_entry);
 717 }
 718
 719 static int ext2_acl_to_xattr(void *dst, const void *src,
 720                              size_t dst_size, size_t src_size)
 721 {
 722         int i, count;
 723         const void *end = src + src_size;
 724         acl_ea_header *ext_acl = (acl_ea_header *)dst;
 725         acl_ea_entry *dst_entry = ext_acl->a_entries;
 726         ext2_acl_entry *src_entry;
 727
 728         if (src_size < sizeof(ext2_acl_header))
 729                 goto fail;
 730         if (((ext2_acl_header *)src)->a_version !=
 731             cpu_to_le32(EXT2_ACL_VERSION))
 732                 goto fail;
 733         src += sizeof(ext2_acl_header);
 734         count = ext2_acl_count(src_size);
 735         if (count <= 0)
 736                 goto fail;
 737
 738         BUG_ON(dst_size < acl_ea_size(count));
 739         ext_acl->a_version = cpu_to_le32(ACL_EA_VERSION);
 740         for (i = 0; i < count; i++, dst_entry++) {
 741                 src_entry = (ext2_acl_entry *)src;
 742                 if (src + sizeof(ext2_acl_entry_short) > end)
 743                         goto fail;
 744                 dst_entry->e_tag = src_entry->e_tag;
 745                 dst_entry->e_perm = src_entry->e_perm;
 746                 switch (le16_to_cpu(src_entry->e_tag)) {
 747                 case ACL_USER_OBJ:
 748                 case ACL_GROUP_OBJ:
 749                 case ACL_MASK:
 750                 case ACL_OTHER:
 751                         src += sizeof(ext2_acl_entry_short);
 752                         dst_entry->e_id = cpu_to_le32(ACL_UNDEFINED_ID);
 753                         break;
 754                 case ACL_USER:
 755                 case ACL_GROUP:
 756                         src += sizeof(ext2_acl_entry);
 757                         if (src > end)
 758                                 goto fail;
 759                         dst_entry->e_id = src_entry->e_id;
 760                         break;
 761                 default:
 762                         goto fail;
 763                 }
 764         }
 765         if (src != end)
 766                 goto fail;
 767         return 0;
 768 fail:
 769         return -EINVAL;
 770 }
 771
 772 static char *xattr_prefix_table[] = {
 773         [1] =   "user.",
 774         [2] =   "system.posix_acl_access",
 775         [3] =   "system.posix_acl_default",
 776         [4] =   "trusted.",
 777         [6] =   "security.",
 778 };
 779
 780 static int copy_single_xattr(struct btrfs_trans_handle *trans,
 781                              struct btrfs_root *root, u64 objectid,
 782                              struct ext2_ext_attr_entry *entry,
 783                              const void *data, u32 datalen)
 784 {
 785         int ret = 0;
 786         int name_len;
 787         int name_index;
 788         void *databuf = NULL;
 789         char namebuf[XATTR_NAME_MAX + 1];
 790
 791         name_index = entry->e_name_index;
 792         if (name_index >= ARRAY_SIZE(xattr_prefix_table) ||
 793             xattr_prefix_table[name_index] == NULL)
 794                 return -EOPNOTSUPP;
 795         name_len = strlen(xattr_prefix_table[name_index]) +
 796                    entry->e_name_len;
 797         if (name_len >= sizeof(namebuf))
 798                 return -ERANGE;
 799
 800         if (name_index == 2 || name_index == 3) {
 801                 size_t bufsize = acl_ea_size(ext2_acl_count(datalen));
 802                 databuf = malloc(bufsize);
 803                 if (!databuf)
 804                        return -ENOMEM;
 805                 ret = ext2_acl_to_xattr(databuf, data, bufsize, datalen);
 806                 if (ret)
 807                         goto out;
 808                 data = databuf;
 809                 datalen = bufsize;
 810         }
 811         strncpy(namebuf, xattr_prefix_table[name_index], XATTR_NAME_MAX);
 812         strncat(namebuf, EXT2_EXT_ATTR_NAME(entry), entry->e_name_len);
 813         if (name_len + datalen > BTRFS_LEAF_DATA_SIZE(root) -
 814             sizeof(struct btrfs_item) - sizeof(struct btrfs_dir_item)) {
 815                 fprintf(stderr, "skip large xattr on inode %Lu name %.*s\n",
 816                         objectid - INO_OFFSET, name_len, namebuf);
 817                 goto out;
 818         }
 819         ret = btrfs_insert_xattr_item(trans, root, namebuf, name_len,
 820                                       data, datalen, objectid);
 821 out:
 822         free(databuf);
 823         return ret;
 824 }
 825
 826 static int copy_extended_attrs(struct btrfs_trans_handle *trans,
 827                                struct btrfs_root *root, u64 objectid,
 828                                struct btrfs_inode_item *btrfs_inode,
 829                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 830 {
 831         int ret = 0;
 832         int inline_ea = 0;
 833         errcode_t err;
 834         u32 datalen;
 835         u32 block_size = ext2_fs->blocksize;
 836         u32 inode_size = EXT2_INODE_SIZE(ext2_fs->super);
 837         struct ext2_inode_large *ext2_inode;
 838         struct ext2_ext_attr_entry *entry;
 839         void *data;
 840         char *buffer = NULL;
 841         char inode_buf[EXT2_GOOD_OLD_INODE_SIZE];
 842
 843         if (inode_size <= EXT2_GOOD_OLD_INODE_SIZE) {
 844                 ext2_inode = (struct ext2_inode_large *)inode_buf;
 845         } else {
 846                 ext2_inode = (struct ext2_inode_large *)malloc(inode_size);
 847                 if (!ext2_inode)
 848                        return -ENOMEM;
 849         }
 850         err = ext2fs_read_inode_full(ext2_fs, ext2_ino, (void *)ext2_inode,
 851                                      inode_size);
 852         if (err) {
 853                 fprintf(stderr, "ext2fs_read_inode_full: %s\n",
 854                         error_message(err));
 855                 ret = -1;
 856                 goto out;
 857         }
 858
 859         if (ext2_ino > ext2_fs->super->s_first_ino &&
 860             inode_size > EXT2_GOOD_OLD_INODE_SIZE) {
 861                 if (EXT2_GOOD_OLD_INODE_SIZE +
 862                     ext2_inode->i_extra_isize > inode_size) {
 863                         ret = -EIO;
 864                         goto out;
 865                 }
 866                 if (ext2_inode->i_extra_isize != 0 &&
 867                     EXT2_XATTR_IHDR(ext2_inode)->h_magic ==
 868                     EXT2_EXT_ATTR_MAGIC) {
 869                         inline_ea = 1;
 870                 }
 871         }
 872         if (inline_ea) {
 873                 int total;
 874                 void *end = (void *)ext2_inode + inode_size;
 875                 entry = EXT2_XATTR_IFIRST(ext2_inode);
 876                 total = end - (void *)entry;
 877                 ret = ext2_xattr_check_names(entry, end);
 878                 if (ret)
 879                         goto out;
 880                 while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 881                         ret = ext2_xattr_check_entry(entry, total);
 882                         if (ret)
 883                                 goto out;
 884                         data = (void *)EXT2_XATTR_IFIRST(ext2_inode) +
 885                                 entry->e_value_offs;
 886                         datalen = entry->e_value_size;
 887                         ret = copy_single_xattr(trans, root, objectid,
 888                                                 entry, data, datalen);
 889                         if (ret)
 890                                 goto out;
 891                         entry = EXT2_EXT_ATTR_NEXT(entry);
 892                 }
 893         }
 894
 895         if (ext2_inode->i_file_acl == 0)
 896                 goto out;
 897
 898         buffer = malloc(block_size);
 899         if (!buffer) {
 900                 ret = -ENOMEM;
 901                 goto out;
 902         }
 903         err = ext2fs_read_ext_attr(ext2_fs, ext2_inode->i_file_acl, buffer);
 904         if (err) {
 905                 fprintf(stderr, "ext2fs_read_ext_attr: %s\n",
 906                         error_message(err));
 907                 ret = -1;
 908                 goto out;
 909         }
 910         ret = ext2_xattr_check_block(buffer, block_size);
 911         if (ret)
 912                 goto out;
 913
 914         entry = EXT2_XATTR_BFIRST(buffer);
 915         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 916                 ret = ext2_xattr_check_entry(entry, block_size);
 917                 if (ret)
 918                         goto out;
 919                 data = buffer + entry->e_value_offs;
 920                 datalen = entry->e_value_size;
 921                 ret = copy_single_xattr(trans, root, objectid,
 922                                         entry, data, datalen);
 923                 if (ret)
 924                         goto out;
 925                 entry = EXT2_EXT_ATTR_NEXT(entry);
 926         }
 927 out:
 928         free(buffer);
 929         if ((void *)ext2_inode != inode_buf)
 930                 free(ext2_inode);
 931         return ret;
 932 }
 933 #define MINORBITS       20
 934 #define MKDEV(ma, mi)   (((ma) << MINORBITS) | (mi))
 935
 936 static inline dev_t old_decode_dev(u16 val)
 937 {
 938         return MKDEV((val >> 8) & 255, val & 255);
 939 }
 940
 941 static inline dev_t new_decode_dev(u32 dev)
 942 {
 943         unsigned major = (dev & 0xfff00) >> 8;
 944         unsigned minor = (dev & 0xff) | ((dev >> 12) & 0xfff00);
 945         return MKDEV(major, minor);
 946 }
 947
 948 static int copy_inode_item(struct btrfs_inode_item *dst,
 949                            struct ext2_inode *src, u32 blocksize)
 950 {
 951         btrfs_set_stack_inode_generation(dst, 1);
 952         btrfs_set_stack_inode_size(dst, src->i_size);
 953         btrfs_set_stack_inode_nbytes(dst, 0);
 954         btrfs_set_stack_inode_block_group(dst, 0);
 955         btrfs_set_stack_inode_nlink(dst, src->i_links_count);
 956         btrfs_set_stack_inode_uid(dst, src->i_uid | (src->i_uid_high << 16));
 957         btrfs_set_stack_inode_gid(dst, src->i_gid | (src->i_gid_high << 16));
 958         btrfs_set_stack_inode_mode(dst, src->i_mode);
 959         btrfs_set_stack_inode_rdev(dst, 0);
 960         btrfs_set_stack_inode_flags(dst, 0);
 961         btrfs_set_stack_timespec_sec(&dst->atime, src->i_atime);
 962         btrfs_set_stack_timespec_nsec(&dst->atime, 0);
 963         btrfs_set_stack_timespec_sec(&dst->ctime, src->i_ctime);
 964         btrfs_set_stack_timespec_nsec(&dst->ctime, 0);
 965         btrfs_set_stack_timespec_sec(&dst->mtime, src->i_mtime);
 966         btrfs_set_stack_timespec_nsec(&dst->mtime, 0);
 967         btrfs_set_stack_timespec_sec(&dst->otime, 0);
 968         btrfs_set_stack_timespec_nsec(&dst->otime, 0);
 969
 970         if (S_ISDIR(src->i_mode)) {
 971                 btrfs_set_stack_inode_size(dst, 0);
 972                 btrfs_set_stack_inode_nlink(dst, 1);
 973         }
 974         if (S_ISREG(src->i_mode)) {
 975                 btrfs_set_stack_inode_size(dst, (u64)src->i_size_high << 32 |
 976                                            (u64)src->i_size);
 977         }
 978         if (!S_ISREG(src->i_mode) && !S_ISDIR(src->i_mode) &&
 979             !S_ISLNK(src->i_mode)) {
 980                 if (src->i_block[0]) {
 981                         btrfs_set_stack_inode_rdev(dst,
 982                                 old_decode_dev(src->i_block[0]));
 983                 } else {
 984                         btrfs_set_stack_inode_rdev(dst,
 985                                 new_decode_dev(src->i_block[1]));
 986                 }
 987         }
 988         return 0;
 989 }
 990
 991 /*
 992  * copy a single inode. do all the required works, such as cloning
 993  * inode item, creating file extents and creating directory entries.
 994  */
 995 static int copy_single_inode(struct btrfs_trans_handle *trans,
 996                              struct btrfs_root *root, u64 objectid,
 997                              ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 998                              struct ext2_inode *ext2_inode,
 999                              int datacsum, int packing, int noxattr)
1000 {
1001         int ret;
1002         struct btrfs_key inode_key;
1003         struct btrfs_inode_item btrfs_inode;
1004
1005         if (ext2_inode->i_links_count == 0)
1006                 return 0;
1007
1008         copy_inode_item(&btrfs_inode, ext2_inode, ext2_fs->blocksize);
1009         if (!datacsum && S_ISREG(ext2_inode->i_mode)) {
1010                 u32 flags = btrfs_stack_inode_flags(&btrfs_inode) |
1011                             BTRFS_INODE_NODATASUM;
1012                 btrfs_set_stack_inode_flags(&btrfs_inode, flags);
1013         }
1014
1015         switch (ext2_inode->i_mode & S_IFMT) {
1016         case S_IFREG:
1017                 ret = create_file_extents(trans, root, objectid, &btrfs_inode,
1018                                         ext2_fs, ext2_ino, datacsum, packing);
1019                 break;
1020         case S_IFDIR:
1021                 ret = create_dir_entries(trans, root, objectid, &btrfs_inode,
1022                                          ext2_fs, ext2_ino);
1023                 break;
1024         case S_IFLNK:
1025                 ret = create_symbol_link(trans, root, objectid, &btrfs_inode,
1026                                          ext2_fs, ext2_ino, ext2_inode);
1027                 break;
1028         default:
1029                 ret = 0;
1030                 break;
1031         }
1032         if (ret)
1033                 return ret;
1034
1035         if (!noxattr) {
1036                 ret = copy_extended_attrs(trans, root, objectid, &btrfs_inode,
1037                                           ext2_fs, ext2_ino);
1038                 if (ret)
1039                         return ret;
1040         }
1041         inode_key.objectid = objectid;
1042         inode_key.offset = 0;
1043         btrfs_set_key_type(&inode_key, BTRFS_INODE_ITEM_KEY);
1044         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1045         return ret;
1046 }
1047
1048 static int copy_disk_extent(struct btrfs_root *root, u64 dst_bytenr,
1049                             u64 src_bytenr, u32 num_bytes)
1050 {
1051         int ret;
1052         char *buffer;
1053         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
1054
1055         buffer = malloc(num_bytes);
1056         if (!buffer)
1057                 return -ENOMEM;
1058         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, src_bytenr);
1059         if (ret != num_bytes)
1060                 goto fail;
1061         ret = pwrite(fs_devs->latest_bdev, buffer, num_bytes, dst_bytenr);
1062         if (ret != num_bytes)
1063                 goto fail;
1064         ret = 0;
1065 fail:
1066         free(buffer);
1067         if (ret > 0)
1068                 ret = -1;
1069         return ret;
1070 }
1071 /*
1072  * scan ext2's inode bitmap and copy all used inodes.
1073  */
1074 static int copy_inodes(struct btrfs_root *root, ext2_filsys ext2_fs,
1075                        int datacsum, int packing, int noxattr, struct task_ctx *p)
1076 {
1077         int ret;
1078         errcode_t err;
1079         ext2_inode_scan ext2_scan;
1080         struct ext2_inode ext2_inode;
1081         ext2_ino_t ext2_ino;
1082         u64 objectid;
1083         struct btrfs_trans_handle *trans;
1084
1085         trans = btrfs_start_transaction(root, 1);
1086         if (!trans)
1087                 return -ENOMEM;
1088         err = ext2fs_open_inode_scan(ext2_fs, 0, &ext2_scan);
1089         if (err) {
1090                 fprintf(stderr, "ext2fs_open_inode_scan: %s\n", error_message(err));
1091                 return -1;
1092         }
1093         while (!(err = ext2fs_get_next_inode(ext2_scan, &ext2_ino,
1094                                              &ext2_inode))) {
1095                 /* no more inodes */
1096                 if (ext2_ino == 0)
1097                         break;
1098                 /* skip special inode in ext2fs */
1099                 if (ext2_ino < EXT2_GOOD_OLD_FIRST_INO &&
1100                     ext2_ino != EXT2_ROOT_INO)
1101                         continue;
1102                 objectid = ext2_ino + INO_OFFSET;
1103                 ret = copy_single_inode(trans, root,
1104                                         objectid, ext2_fs, ext2_ino,
1105                                         &ext2_inode, datacsum, packing,
1106                                         noxattr);
1107                 p->cur_copy_inodes++;
1108                 if (ret)
1109                         return ret;
1110                 if (trans->blocks_used >= 4096) {
1111                         ret = btrfs_commit_transaction(trans, root);
1112                         BUG_ON(ret);
1113                         trans = btrfs_start_transaction(root, 1);
1114                         BUG_ON(!trans);
1115                 }
1116         }
1117         if (err) {
1118                 fprintf(stderr, "ext2fs_get_next_inode: %s\n", error_message(err));
1119                 return -1;
1120         }
1121         ret = btrfs_commit_transaction(trans, root);
1122         BUG_ON(ret);
1123
1124         return ret;
1125 }
1126
1127 /*
1128  * Construct a range of ext2fs image file.
1129  * scan block allocation bitmap, find all blocks used by the ext2fs
1130  * in this range and create file extents that point to these blocks.
1131  *
1132  * Note: Before calling the function, no file extent points to blocks
1133  *       in this range
1134  */
1135 static int create_image_file_range(struct btrfs_trans_handle *trans,
1136                                    struct btrfs_root *root, u64 objectid,
1137                                    struct btrfs_inode_item *inode,
1138                                    u64 start_byte, u64 end_byte,
1139                                    ext2_filsys ext2_fs)
1140 {
1141         u32 blocksize = ext2_fs->blocksize;
1142         u32 block = start_byte / blocksize;
1143         u32 last_block = (end_byte + blocksize - 1) / blocksize;
1144         int ret = 0;
1145         struct blk_iterate_data data = {
1146                 .trans          = trans,
1147                 .root           = root,
1148                 .inode          = inode,
1149                 .objectid       = objectid,
1150                 .first_block    = block,
1151                 .disk_block     = 0,
1152                 .num_blocks     = 0,
1153                 .boundary       = (u64)-1,
1154                 .checksum       = 0,
1155                 .errcode        = 0,
1156         };
1157         for (; start_byte < end_byte; block++, start_byte += blocksize) {
1158                 if (!ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
1159                         continue;
1160                 ret = block_iterate_proc(NULL, block, block, &data);
1161                 if (ret & BLOCK_ABORT) {
1162                         ret = data.errcode;
1163                         goto fail;
1164                 }
1165         }
1166         if (data.num_blocks > 0) {
1167                 ret = record_file_blocks(trans, root, objectid, inode,
1168                                          data.first_block, data.disk_block,
1169                                          data.num_blocks, 0);
1170                 if (ret)
1171                         goto fail;
1172                 data.first_block += data.num_blocks;
1173         }
1174         if (last_block > data.first_block) {
1175                 ret = record_file_blocks(trans, root, objectid, inode,
1176                                          data.first_block, 0, last_block -
1177                                          data.first_block, 0);
1178                 if (ret)
1179                         goto fail;
1180         }
1181 fail:
1182         return ret;
1183 }
1184 /*
1185  * Create the ext2fs image file.
1186  */
1187 static int create_ext2_image(struct btrfs_root *root, ext2_filsys ext2_fs,
1188                              const char *name)
1189 {
1190         int ret;
1191         struct btrfs_key key;
1192         struct btrfs_key location;
1193         struct btrfs_path path;
1194         struct btrfs_inode_item btrfs_inode;
1195         struct btrfs_inode_item *inode_item;
1196         struct extent_buffer *leaf;
1197         struct btrfs_fs_info *fs_info = root->fs_info;
1198         struct btrfs_root *extent_root = fs_info->extent_root;
1199         struct btrfs_trans_handle *trans;
1200         struct btrfs_extent_item *ei;
1201         struct btrfs_extent_inline_ref *iref;
1202         struct btrfs_extent_data_ref *dref;
1203         u64 bytenr;
1204         u64 num_bytes;
1205         u64 objectid;
1206         u64 last_byte;
1207         u64 first_free;
1208         u64 total_bytes;
1209         u32 sectorsize = root->sectorsize;
1210
1211         total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
1212         first_free =  BTRFS_SUPER_INFO_OFFSET + sectorsize * 2 - 1;
1213         first_free &= ~((u64)sectorsize - 1);
1214
1215         memset(&btrfs_inode, 0, sizeof(btrfs_inode));
1216         btrfs_set_stack_inode_generation(&btrfs_inode, 1);
1217         btrfs_set_stack_inode_size(&btrfs_inode, total_bytes);
1218         btrfs_set_stack_inode_nlink(&btrfs_inode, 1);
1219         btrfs_set_stack_inode_nbytes(&btrfs_inode, 0);
1220         btrfs_set_stack_inode_mode(&btrfs_inode, S_IFREG | 0400);
1221         btrfs_set_stack_inode_flags(&btrfs_inode, BTRFS_INODE_NODATASUM |
1222                                     BTRFS_INODE_READONLY);
1223         btrfs_init_path(&path);
1224         trans = btrfs_start_transaction(root, 1);
1225         BUG_ON(!trans);
1226
1227         objectid = btrfs_root_dirid(&root->root_item);
1228         ret = btrfs_find_free_objectid(trans, root, objectid, &objectid);
1229         if (ret)
1230                 goto fail;
1231
1232         /*
1233          * copy blocks covered by extent #0 to new positions. extent #0 is
1234          * special, we can't rely on relocate_extents_range to relocate it.
1235          */
1236         for (last_byte = 0; last_byte < first_free; last_byte += sectorsize) {
1237                 ret = custom_alloc_extent(root, sectorsize, 0, &key);
1238                 if (ret)
1239                         goto fail;
1240                 ret = copy_disk_extent(root, key.objectid, last_byte,
1241                                        sectorsize);
1242                 if (ret)
1243                         goto fail;
1244                 ret = btrfs_record_file_extent(trans, root, objectid,
1245                                                &btrfs_inode, last_byte,
1246                                                key.objectid, sectorsize);
1247                 if (ret)
1248                         goto fail;
1249         }
1250
1251         while(1) {
1252                 key.objectid = last_byte;
1253                 key.offset = 0;
1254                 btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
1255                 ret = btrfs_search_slot(trans, fs_info->extent_root,
1256                                         &key, &path, 0, 0);
1257                 if (ret < 0)
1258                         goto fail;
1259 next:
1260                 leaf = path.nodes[0];
1261                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1262                         ret = btrfs_next_leaf(extent_root, &path);
1263                         if (ret < 0)
1264                                 goto fail;
1265                         if (ret > 0)
1266                                 break;
1267                         leaf = path.nodes[0];
1268                 }
1269                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1270                 if (last_byte > key.objectid ||
1271                     key.type != BTRFS_EXTENT_ITEM_KEY) {
1272                         path.slots[0]++;
1273                         goto next;
1274                 }
1275
1276                 bytenr = key.objectid;
1277                 num_bytes = key.offset;
1278                 ei = btrfs_item_ptr(leaf, path.slots[0],
1279                                     struct btrfs_extent_item);
1280                 if (!(btrfs_extent_flags(leaf, ei) & BTRFS_EXTENT_FLAG_DATA)) {
1281                         path.slots[0]++;
1282                         goto next;
1283                 }
1284
1285                 BUG_ON(btrfs_item_size_nr(leaf, path.slots[0]) != sizeof(*ei) +
1286                        btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY));
1287
1288                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
1289                 key.type = btrfs_extent_inline_ref_type(leaf, iref);
1290                 BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
1291                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1292                 if (btrfs_extent_data_ref_root(leaf, dref) !=
1293                     BTRFS_FS_TREE_OBJECTID) {
1294                         path.slots[0]++;
1295                         goto next;
1296                 }
1297
1298                 if (bytenr > last_byte) {
1299                         ret = create_image_file_range(trans, root, objectid,
1300                                                       &btrfs_inode, last_byte,
1301                                                       bytenr, ext2_fs);
1302                         if (ret)
1303                                 goto fail;
1304                 }
1305                 ret = btrfs_record_file_extent(trans, root, objectid,
1306                                                &btrfs_inode, bytenr, bytenr,
1307                                                num_bytes);
1308                 if (ret)
1309                         goto fail;
1310                 last_byte = bytenr + num_bytes;
1311                 btrfs_release_path(&path);
1312
1313                 if (trans->blocks_used >= 4096) {
1314                         ret = btrfs_commit_transaction(trans, root);
1315                         BUG_ON(ret);
1316                         trans = btrfs_start_transaction(root, 1);
1317                         BUG_ON(!trans);
1318                 }
1319         }
1320         btrfs_release_path(&path);
1321         if (total_bytes > last_byte) {
1322                 ret = create_image_file_range(trans, root, objectid,
1323                                               &btrfs_inode, last_byte,
1324                                               total_bytes, ext2_fs);
1325                 if (ret)
1326                         goto fail;
1327         }
1328
1329         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1330         if (ret)
1331                 goto fail;
1332
1333         location.objectid = objectid;
1334         location.offset = 0;
1335         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1336         ret = btrfs_insert_dir_item(trans, root, name, strlen(name),
1337                                     btrfs_root_dirid(&root->root_item),
1338                                     &location, EXT2_FT_REG_FILE, objectid);
1339         if (ret)
1340                 goto fail;
1341         ret = btrfs_insert_inode_ref(trans, root, name, strlen(name),
1342                                      objectid,
1343                                      btrfs_root_dirid(&root->root_item),
1344                                      objectid);
1345         if (ret)
1346                 goto fail;
1347         location.objectid = btrfs_root_dirid(&root->root_item);
1348         location.offset = 0;
1349         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1350         ret = btrfs_lookup_inode(trans, root, &path, &location, 1);
1351         if (ret)
1352                 goto fail;
1353         leaf = path.nodes[0];
1354         inode_item = btrfs_item_ptr(leaf, path.slots[0],
1355                                     struct btrfs_inode_item);
1356         btrfs_set_inode_size(leaf, inode_item, strlen(name) * 2 +
1357                              btrfs_inode_size(leaf, inode_item));
1358         btrfs_mark_buffer_dirty(leaf);
1359         btrfs_release_path(&path);
1360         ret = btrfs_commit_transaction(trans, root);
1361         BUG_ON(ret);
1362 fail:
1363         btrfs_release_path(&path);
1364         return ret;
1365 }
1366
1367 static struct btrfs_root * link_subvol(struct btrfs_root *root,
1368                 const char *base, u64 root_objectid)
1369 {
1370         struct btrfs_trans_handle *trans;
1371         struct btrfs_fs_info *fs_info = root->fs_info;
1372         struct btrfs_root *tree_root = fs_info->tree_root;
1373         struct btrfs_root *new_root = NULL;
1374         struct btrfs_path *path;
1375         struct btrfs_inode_item *inode_item;
1376         struct extent_buffer *leaf;
1377         struct btrfs_key key;
1378         u64 dirid = btrfs_root_dirid(&root->root_item);
1379         u64 index = 2;
1380         char buf[BTRFS_NAME_LEN + 1]; /* for snprintf null */
1381         int len;
1382         int i;
1383         int ret;
1384
1385         len = strlen(base);
1386         if (len < 1 || len > BTRFS_NAME_LEN)
1387                 return NULL;
1388
1389         path = btrfs_alloc_path();
1390         BUG_ON(!path);
1391
1392         key.objectid = dirid;
1393         key.type = BTRFS_DIR_INDEX_KEY;
1394         key.offset = (u64)-1;
1395
1396         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1397         BUG_ON(ret <= 0);
1398
1399         if (path->slots[0] > 0) {
1400                 path->slots[0]--;
1401                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1402                 if (key.objectid == dirid && key.type == BTRFS_DIR_INDEX_KEY)
1403                         index = key.offset + 1;
1404         }
1405         btrfs_release_path(path);
1406
1407         trans = btrfs_start_transaction(root, 1);
1408         BUG_ON(!trans);
1409
1410         key.objectid = dirid;
1411         key.offset = 0;
1412         key.type =  BTRFS_INODE_ITEM_KEY;
1413
1414         ret = btrfs_lookup_inode(trans, root, path, &key, 1);
1415         BUG_ON(ret);
1416         leaf = path->nodes[0];
1417         inode_item = btrfs_item_ptr(leaf, path->slots[0],
1418                                     struct btrfs_inode_item);
1419
1420         key.objectid = root_objectid;
1421         key.offset = (u64)-1;
1422         key.type = BTRFS_ROOT_ITEM_KEY;
1423
1424         memcpy(buf, base, len);
1425         for (i = 0; i < 1024; i++) {
1426                 ret = btrfs_insert_dir_item(trans, root, buf, len,
1427                                             dirid, &key, BTRFS_FT_DIR, index);
1428                 if (ret != -EEXIST)
1429                         break;
1430                 len = snprintf(buf, ARRAY_SIZE(buf), "%s%d", base, i);
1431                 if (len < 1 || len > BTRFS_NAME_LEN) {
1432                         ret = -EINVAL;
1433                         break;
1434                 }
1435         }
1436         if (ret)
1437                 goto fail;
1438
1439         btrfs_set_inode_size(leaf, inode_item, len * 2 +
1440                              btrfs_inode_size(leaf, inode_item));
1441         btrfs_mark_buffer_dirty(leaf);
1442         btrfs_release_path(path);
1443
1444         /* add the backref first */
1445         ret = btrfs_add_root_ref(trans, tree_root, root_objectid,
1446                                  BTRFS_ROOT_BACKREF_KEY,
1447                                  root->root_key.objectid,
1448                                  dirid, index, buf, len);
1449         BUG_ON(ret);
1450
1451         /* now add the forward ref */
1452         ret = btrfs_add_root_ref(trans, tree_root, root->root_key.objectid,
1453                                  BTRFS_ROOT_REF_KEY, root_objectid,
1454                                  dirid, index, buf, len);
1455
1456         ret = btrfs_commit_transaction(trans, root);
1457         BUG_ON(ret);
1458
1459         new_root = btrfs_read_fs_root(fs_info, &key);
1460         if (IS_ERR(new_root))
1461                 new_root = NULL;
1462 fail:
1463         btrfs_free_path(path);
1464         return new_root;
1465 }
1466
1467 static int create_chunk_mapping(struct btrfs_trans_handle *trans,
1468                                 struct btrfs_root *root)
1469 {
1470         struct btrfs_fs_info *info = root->fs_info;
1471         struct btrfs_root *chunk_root = info->chunk_root;
1472         struct btrfs_root *extent_root = info->extent_root;
1473         struct btrfs_device *device;
1474         struct btrfs_block_group_cache *cache;
1475         struct btrfs_dev_extent *extent;
1476         struct extent_buffer *leaf;
1477         struct btrfs_chunk chunk;
1478         struct btrfs_key key;
1479         struct btrfs_path path;
1480         u64 cur_start;
1481         u64 total_bytes;
1482         u64 chunk_objectid;
1483         int ret;
1484
1485         btrfs_init_path(&path);
1486
1487         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
1488         chunk_objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
1489
1490         BUG_ON(list_empty(&info->fs_devices->devices));
1491         device = list_entry(info->fs_devices->devices.next,
1492                             struct btrfs_device, dev_list);
1493         BUG_ON(device->devid != info->fs_devices->latest_devid);
1494
1495         /* delete device extent created by make_btrfs */
1496         key.objectid = device->devid;
1497         key.offset = 0;
1498         key.type = BTRFS_DEV_EXTENT_KEY;
1499         ret = btrfs_search_slot(trans, device->dev_root, &key, &path, -1, 1);
1500         if (ret < 0)
1501                 goto err;
1502
1503         BUG_ON(ret > 0);
1504         ret = btrfs_del_item(trans, device->dev_root, &path);
1505         if (ret)
1506                 goto err;
1507         btrfs_release_path(&path);
1508
1509         /* delete chunk item created by make_btrfs */
1510         key.objectid = chunk_objectid;
1511         key.offset = 0;
1512         key.type = BTRFS_CHUNK_ITEM_KEY;
1513         ret = btrfs_search_slot(trans, chunk_root, &key, &path, -1, 1);
1514         if (ret < 0)
1515                 goto err;
1516
1517         BUG_ON(ret > 0);
1518         ret = btrfs_del_item(trans, chunk_root, &path);
1519         if (ret)
1520                 goto err;
1521         btrfs_release_path(&path);
1522
1523         /* for each block group, create device extent and chunk item */
1524         cur_start = 0;
1525         while (cur_start < total_bytes) {
1526                 cache = btrfs_lookup_block_group(root->fs_info, cur_start);
1527                 BUG_ON(!cache);
1528
1529                 /* insert device extent */
1530                 key.objectid = device->devid;
1531                 key.offset = cache->key.objectid;
1532                 key.type = BTRFS_DEV_EXTENT_KEY;
1533                 ret = btrfs_insert_empty_item(trans, device->dev_root, &path,
1534                                               &key, sizeof(*extent));
1535                 if (ret)
1536                         goto err;
1537
1538                 leaf = path.nodes[0];
1539                 extent = btrfs_item_ptr(leaf, path.slots[0],
1540                                         struct btrfs_dev_extent);
1541
1542                 btrfs_set_dev_extent_chunk_tree(leaf, extent,
1543                                                 chunk_root->root_key.objectid);
1544                 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
1545                                                     chunk_objectid);
1546                 btrfs_set_dev_extent_chunk_offset(leaf, extent,
1547                                                   cache->key.objectid);
1548                 btrfs_set_dev_extent_length(leaf, extent, cache->key.offset);
1549                 write_extent_buffer(leaf, root->fs_info->chunk_tree_uuid,
1550                     (unsigned long)btrfs_dev_extent_chunk_tree_uuid(extent),
1551                     BTRFS_UUID_SIZE);
1552                 btrfs_mark_buffer_dirty(leaf);
1553                 btrfs_release_path(&path);
1554
1555                 /* insert chunk item */
1556                 btrfs_set_stack_chunk_length(&chunk, cache->key.offset);
1557                 btrfs_set_stack_chunk_owner(&chunk,
1558                                             extent_root->root_key.objectid);
1559                 btrfs_set_stack_chunk_stripe_len(&chunk, BTRFS_STRIPE_LEN);
1560                 btrfs_set_stack_chunk_type(&chunk, cache->flags);
1561                 btrfs_set_stack_chunk_io_align(&chunk, device->io_align);
1562                 btrfs_set_stack_chunk_io_width(&chunk, device->io_width);
1563                 btrfs_set_stack_chunk_sector_size(&chunk, device->sector_size);
1564                 btrfs_set_stack_chunk_num_stripes(&chunk, 1);
1565                 btrfs_set_stack_chunk_sub_stripes(&chunk, 0);
1566                 btrfs_set_stack_stripe_devid(&chunk.stripe, device->devid);
1567                 btrfs_set_stack_stripe_offset(&chunk.stripe,
1568                                               cache->key.objectid);
1569                 memcpy(&chunk.stripe.dev_uuid, device->uuid, BTRFS_UUID_SIZE);
1570
1571                 key.objectid = chunk_objectid;
1572                 key.offset = cache->key.objectid;
1573                 key.type = BTRFS_CHUNK_ITEM_KEY;
1574
1575                 ret = btrfs_insert_item(trans, chunk_root, &key, &chunk,
1576                                         btrfs_chunk_item_size(1));
1577                 if (ret)
1578                         goto err;
1579
1580                 cur_start = cache->key.objectid + cache->key.offset;
1581         }
1582
1583         device->bytes_used = total_bytes;
1584         ret = btrfs_update_device(trans, device);
1585 err:
1586         btrfs_release_path(&path);
1587         return ret;
1588 }
1589
1590 static int create_subvol(struct btrfs_trans_handle *trans,
1591                          struct btrfs_root *root, u64 root_objectid)
1592 {
1593         struct extent_buffer *tmp;
1594         struct btrfs_root *new_root;
1595         struct btrfs_key key;
1596         struct btrfs_root_item root_item;
1597         int ret;
1598
1599         ret = btrfs_copy_root(trans, root, root->node, &tmp,
1600                               root_objectid);
1601         BUG_ON(ret);
1602
1603         memcpy(&root_item, &root->root_item, sizeof(root_item));
1604         btrfs_set_root_bytenr(&root_item, tmp->start);
1605         btrfs_set_root_level(&root_item, btrfs_header_level(tmp));
1606         btrfs_set_root_generation(&root_item, trans->transid);
1607         free_extent_buffer(tmp);
1608
1609         key.objectid = root_objectid;
1610         key.type = BTRFS_ROOT_ITEM_KEY;
1611         key.offset = trans->transid;
1612         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1613                                 &key, &root_item);
1614
1615         key.offset = (u64)-1;
1616         new_root = btrfs_read_fs_root(root->fs_info, &key);
1617         BUG_ON(!new_root || IS_ERR(new_root));
1618
1619         ret = btrfs_make_root_dir(trans, new_root, BTRFS_FIRST_FREE_OBJECTID);
1620         BUG_ON(ret);
1621
1622         return 0;
1623 }
1624
1625 static int init_btrfs(struct btrfs_root *root)
1626 {
1627         int ret;
1628         struct btrfs_key location;
1629         struct btrfs_trans_handle *trans;
1630         struct btrfs_fs_info *fs_info = root->fs_info;
1631         struct extent_buffer *tmp;
1632
1633         trans = btrfs_start_transaction(root, 1);
1634         BUG_ON(!trans);
1635         ret = btrfs_make_block_groups(trans, root);
1636         if (ret)
1637                 goto err;
1638         ret = btrfs_fix_block_accounting(trans, root);
1639         if (ret)
1640                 goto err;
1641         ret = create_chunk_mapping(trans, root);
1642         if (ret)
1643                 goto err;
1644         ret = btrfs_make_root_dir(trans, fs_info->tree_root,
1645                                   BTRFS_ROOT_TREE_DIR_OBJECTID);
1646         if (ret)
1647                 goto err;
1648         memcpy(&location, &root->root_key, sizeof(location));
1649         location.offset = (u64)-1;
1650         ret = btrfs_insert_dir_item(trans, fs_info->tree_root, "default", 7,
1651                                 btrfs_super_root_dir(fs_info->super_copy),
1652                                 &location, BTRFS_FT_DIR, 0);
1653         if (ret)
1654                 goto err;
1655         ret = btrfs_insert_inode_ref(trans, fs_info->tree_root, "default", 7,
1656                                 location.objectid,
1657                                 btrfs_super_root_dir(fs_info->super_copy), 0);
1658         if (ret)
1659                 goto err;
1660         btrfs_set_root_dirid(&fs_info->fs_root->root_item,
1661                              BTRFS_FIRST_FREE_OBJECTID);
1662
1663         /* subvol for ext2 image file */
1664         ret = create_subvol(trans, root, EXT2_IMAGE_SUBVOL_OBJECTID);
1665         BUG_ON(ret);
1666         /* subvol for data relocation */
1667         ret = create_subvol(trans, root, BTRFS_DATA_RELOC_TREE_OBJECTID);
1668         BUG_ON(ret);
1669
1670         extent_buffer_get(fs_info->csum_root->node);
1671         ret = __btrfs_cow_block(trans, fs_info->csum_root,
1672                                 fs_info->csum_root->node, NULL, 0, &tmp, 0, 0);
1673         BUG_ON(ret);
1674         free_extent_buffer(tmp);
1675
1676         ret = btrfs_commit_transaction(trans, root);
1677         BUG_ON(ret);
1678 err:
1679         return ret;
1680 }
1681
1682 /*
1683  * Migrate super block to its default position and zero 0 ~ 16k
1684  */
1685 static int migrate_super_block(int fd, u64 old_bytenr, u32 sectorsize)
1686 {
1687         int ret;
1688         struct extent_buffer *buf;
1689         struct btrfs_super_block *super;
1690         u32 len;
1691         u32 bytenr;
1692
1693         BUG_ON(sectorsize < sizeof(*super));
1694         buf = malloc(sizeof(*buf) + sectorsize);
1695         if (!buf)
1696                 return -ENOMEM;
1697
1698         buf->len = sectorsize;
1699         ret = pread(fd, buf->data, sectorsize, old_bytenr);
1700         if (ret != sectorsize)
1701                 goto fail;
1702
1703         super = (struct btrfs_super_block *)buf->data;
1704         BUG_ON(btrfs_super_bytenr(super) != old_bytenr);
1705         btrfs_set_super_bytenr(super, BTRFS_SUPER_INFO_OFFSET);
1706
1707         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1708         ret = pwrite(fd, buf->data, sectorsize, BTRFS_SUPER_INFO_OFFSET);
1709         if (ret != sectorsize)
1710                 goto fail;
1711
1712         ret = fsync(fd);
1713         if (ret)
1714                 goto fail;
1715
1716         memset(buf->data, 0, sectorsize);
1717         for (bytenr = 0; bytenr < BTRFS_SUPER_INFO_OFFSET; ) {
1718                 len = BTRFS_SUPER_INFO_OFFSET - bytenr;
1719                 if (len > sectorsize)
1720                         len = sectorsize;
1721                 ret = pwrite(fd, buf->data, len, bytenr);
1722                 if (ret != len) {
1723                         fprintf(stderr, "unable to zero fill device\n");
1724                         break;
1725                 }
1726                 bytenr += len;
1727         }
1728         ret = 0;
1729         fsync(fd);
1730 fail:
1731         free(buf);
1732         if (ret > 0)
1733                 ret = -1;
1734         return ret;
1735 }
1736
1737 static int prepare_system_chunk_sb(struct btrfs_super_block *super)
1738 {
1739         struct btrfs_chunk *chunk;
1740         struct btrfs_disk_key *key;
1741         u32 sectorsize = btrfs_super_sectorsize(super);
1742
1743         key = (struct btrfs_disk_key *)(super->sys_chunk_array);
1744         chunk = (struct btrfs_chunk *)(super->sys_chunk_array +
1745                                        sizeof(struct btrfs_disk_key));
1746
1747         btrfs_set_disk_key_objectid(key, BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1748         btrfs_set_disk_key_type(key, BTRFS_CHUNK_ITEM_KEY);
1749         btrfs_set_disk_key_offset(key, 0);
1750
1751         btrfs_set_stack_chunk_length(chunk, btrfs_super_total_bytes(super));
1752         btrfs_set_stack_chunk_owner(chunk, BTRFS_EXTENT_TREE_OBJECTID);
1753         btrfs_set_stack_chunk_stripe_len(chunk, BTRFS_STRIPE_LEN);
1754         btrfs_set_stack_chunk_type(chunk, BTRFS_BLOCK_GROUP_SYSTEM);
1755         btrfs_set_stack_chunk_io_align(chunk, sectorsize);
1756         btrfs_set_stack_chunk_io_width(chunk, sectorsize);
1757         btrfs_set_stack_chunk_sector_size(chunk, sectorsize);
1758         btrfs_set_stack_chunk_num_stripes(chunk, 1);
1759         btrfs_set_stack_chunk_sub_stripes(chunk, 0);
1760         chunk->stripe.devid = super->dev_item.devid;
1761         btrfs_set_stack_stripe_offset(&chunk->stripe, 0);
1762         memcpy(chunk->stripe.dev_uuid, super->dev_item.uuid, BTRFS_UUID_SIZE);
1763         btrfs_set_super_sys_array_size(super, sizeof(*key) + sizeof(*chunk));
1764         return 0;
1765 }
1766
1767 static int prepare_system_chunk(int fd, u64 sb_bytenr, u32 sectorsize)
1768 {
1769         int ret;
1770         struct extent_buffer *buf;
1771         struct btrfs_super_block *super;
1772
1773         BUG_ON(sectorsize < sizeof(*super));
1774         buf = malloc(sizeof(*buf) + sectorsize);
1775         if (!buf)
1776                 return -ENOMEM;
1777
1778         buf->len = sectorsize;
1779         ret = pread(fd, buf->data, sectorsize, sb_bytenr);
1780         if (ret != sectorsize)
1781                 goto fail;
1782
1783         super = (struct btrfs_super_block *)buf->data;
1784         BUG_ON(btrfs_super_bytenr(super) != sb_bytenr);
1785         BUG_ON(btrfs_super_num_devices(super) != 1);
1786
1787         ret = prepare_system_chunk_sb(super);
1788         if (ret)
1789                 goto fail;
1790
1791         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1792         ret = pwrite(fd, buf->data, sectorsize, sb_bytenr);
1793         if (ret != sectorsize)
1794                 goto fail;
1795
1796         ret = 0;
1797 fail:
1798         free(buf);
1799         if (ret > 0)
1800                 ret = -1;
1801         return ret;
1802 }
1803
1804 static int relocate_one_reference(struct btrfs_trans_handle *trans,
1805                                   struct btrfs_root *root,
1806                                   u64 extent_start, u64 extent_size,
1807                                   struct btrfs_key *extent_key,
1808                                   struct extent_io_tree *reloc_tree)
1809 {
1810         struct extent_buffer *leaf;
1811         struct btrfs_file_extent_item *fi;
1812         struct btrfs_key key;
1813         struct btrfs_path path;
1814         struct btrfs_inode_item inode;
1815         struct blk_iterate_data data;
1816         u64 bytenr;
1817         u64 num_bytes;
1818         u64 cur_offset;
1819         u64 new_pos;
1820         u64 nbytes;
1821         u64 sector_end;
1822         u32 sectorsize = root->sectorsize;
1823         unsigned long ptr;
1824         int datacsum;
1825         int fd;
1826         int ret;
1827
1828         btrfs_init_path(&path);
1829         ret = btrfs_search_slot(trans, root, extent_key, &path, -1, 1);
1830         if (ret)
1831                 goto fail;
1832
1833         leaf = path.nodes[0];
1834         fi = btrfs_item_ptr(leaf, path.slots[0],
1835                             struct btrfs_file_extent_item);
1836         BUG_ON(btrfs_file_extent_offset(leaf, fi) > 0);
1837         if (extent_start != btrfs_file_extent_disk_bytenr(leaf, fi) ||
1838             extent_size != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1839                 ret = 1;
1840                 goto fail;
1841         }
1842
1843         bytenr = extent_start + btrfs_file_extent_offset(leaf, fi);
1844         num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
1845
1846         ret = btrfs_del_item(trans, root, &path);
1847         if (ret)
1848                 goto fail;
1849
1850         ret = btrfs_free_extent(trans, root, extent_start, extent_size, 0,
1851                                 root->root_key.objectid,
1852                                 extent_key->objectid, extent_key->offset);
1853         if (ret)
1854                 goto fail;
1855
1856         btrfs_release_path(&path);
1857
1858         key.objectid = extent_key->objectid;
1859         key.offset = 0;
1860         key.type =  BTRFS_INODE_ITEM_KEY;
1861         ret = btrfs_lookup_inode(trans, root, &path, &key, 0);
1862         if (ret)
1863                 goto fail;
1864
1865         leaf = path.nodes[0];
1866         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1867         read_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1868         btrfs_release_path(&path);
1869
1870         BUG_ON(num_bytes & (sectorsize - 1));
1871         nbytes = btrfs_stack_inode_nbytes(&inode) - num_bytes;
1872         btrfs_set_stack_inode_nbytes(&inode, nbytes);
1873         datacsum = !(btrfs_stack_inode_flags(&inode) & BTRFS_INODE_NODATASUM);
1874
1875         data = (struct blk_iterate_data) {
1876                 .trans          = trans,
1877                 .root           = root,
1878                 .inode          = &inode,
1879                 .objectid       = extent_key->objectid,
1880                 .first_block    = extent_key->offset / sectorsize,
1881                 .disk_block     = 0,
1882                 .num_blocks     = 0,
1883                 .boundary       = (u64)-1,
1884                 .checksum       = datacsum,
1885                 .errcode        = 0,
1886         };
1887
1888         cur_offset = extent_key->offset;
1889         while (num_bytes > 0) {
1890                 sector_end = bytenr + sectorsize - 1;
1891                 if (test_range_bit(reloc_tree, bytenr, sector_end,
1892                                    EXTENT_LOCKED, 1)) {
1893                         ret = get_state_private(reloc_tree, bytenr, &new_pos);
1894                         BUG_ON(ret);
1895                 } else {
1896                         ret = custom_alloc_extent(root, sectorsize, 0, &key);
1897                         if (ret)
1898                                 goto fail;
1899                         new_pos = key.objectid;
1900
1901                         if (cur_offset == extent_key->offset) {
1902                                 fd = root->fs_info->fs_devices->latest_bdev;
1903                                 readahead(fd, bytenr, num_bytes);
1904                         }
1905                         ret = copy_disk_extent(root, new_pos, bytenr,
1906                                                sectorsize);
1907                         if (ret)
1908                                 goto fail;
1909                         ret = set_extent_bits(reloc_tree, bytenr, sector_end,
1910                                               EXTENT_LOCKED, GFP_NOFS);
1911                         BUG_ON(ret);
1912                         ret = set_state_private(reloc_tree, bytenr, new_pos);
1913                         BUG_ON(ret);
1914                 }
1915
1916                 ret = block_iterate_proc(NULL, new_pos / sectorsize,
1917                                          cur_offset / sectorsize, &data);
1918                 if (ret & BLOCK_ABORT) {
1919                         ret = data.errcode;
1920                         goto fail;
1921                 }
1922
1923                 cur_offset += sectorsize;
1924                 bytenr += sectorsize;
1925                 num_bytes -= sectorsize;
1926         }
1927
1928         if (data.num_blocks > 0) {
1929                 ret = record_file_blocks(trans, root,
1930                                          extent_key->objectid, &inode,
1931                                          data.first_block, data.disk_block,
1932                                          data.num_blocks, datacsum);
1933                 if (ret)
1934                         goto fail;
1935         }
1936
1937         key.objectid = extent_key->objectid;
1938         key.offset = 0;
1939         key.type =  BTRFS_INODE_ITEM_KEY;
1940         ret = btrfs_lookup_inode(trans, root, &path, &key, 1);
1941         if (ret)
1942                 goto fail;
1943
1944         leaf = path.nodes[0];
1945         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1946         write_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1947         btrfs_mark_buffer_dirty(leaf);
1948         btrfs_release_path(&path);
1949
1950 fail:
1951         btrfs_release_path(&path);
1952         return ret;
1953 }
1954
1955 static int relocate_extents_range(struct btrfs_root *fs_root,
1956                                   struct btrfs_root *ext2_root,
1957                                   u64 start_byte, u64 end_byte)
1958 {
1959         struct btrfs_fs_info *info = fs_root->fs_info;
1960         struct btrfs_root *extent_root = info->extent_root;
1961         struct btrfs_root *cur_root = NULL;
1962         struct btrfs_trans_handle *trans;
1963         struct btrfs_extent_data_ref *dref;
1964         struct btrfs_extent_inline_ref *iref;
1965         struct btrfs_extent_item *ei;
1966         struct extent_buffer *leaf;
1967         struct btrfs_key key;
1968         struct btrfs_key extent_key;
1969         struct btrfs_path path;
1970         struct extent_io_tree reloc_tree;
1971         unsigned long ptr;
1972         unsigned long end;
1973         u64 cur_byte;
1974         u64 num_bytes;
1975         u64 ref_root;
1976         u64 num_extents;
1977         int pass = 0;
1978         int ret;
1979
1980         btrfs_init_path(&path);
1981         extent_io_tree_init(&reloc_tree);
1982
1983         key.objectid = start_byte;
1984         key.offset = 0;
1985         key.type = BTRFS_EXTENT_ITEM_KEY;
1986         ret = btrfs_search_slot(NULL, extent_root, &key, &path, 0, 0);
1987         if (ret < 0)
1988                 goto fail;
1989         if (ret > 0) {
1990                 ret = btrfs_previous_item(extent_root, &path, 0,
1991                                           BTRFS_EXTENT_ITEM_KEY);
1992                 if (ret < 0)
1993                         goto fail;
1994                 if (ret == 0) {
1995                         leaf = path.nodes[0];
1996                         btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1997                         if (key.objectid + key.offset > start_byte)
1998                                 start_byte = key.objectid;
1999                 }
2000         }
2001         btrfs_release_path(&path);
2002 again:
2003         cur_root = (pass % 2 == 0) ? ext2_root : fs_root;
2004         num_extents = 0;
2005
2006         trans = btrfs_start_transaction(cur_root, 1);
2007         BUG_ON(!trans);
2008
2009         cur_byte = start_byte;
2010         while (1) {
2011                 key.objectid = cur_byte;
2012                 key.offset = 0;
2013                 key.type = BTRFS_EXTENT_ITEM_KEY;
2014                 ret = btrfs_search_slot(trans, extent_root,
2015                                         &key, &path, 0, 0);
2016                 if (ret < 0)
2017                         goto fail;
2018 next:
2019                 leaf = path.nodes[0];
2020                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2021                         ret = btrfs_next_leaf(extent_root, &path);
2022                         if (ret < 0)
2023                                 goto fail;
2024                         if (ret > 0)
2025                                 break;
2026                         leaf = path.nodes[0];
2027                 }
2028
2029                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2030                 if (key.objectid < cur_byte ||
2031                     key.type != BTRFS_EXTENT_ITEM_KEY) {
2032                         path.slots[0]++;
2033                         goto next;
2034                 }
2035                 if (key.objectid >= end_byte)
2036                         break;
2037
2038                 num_extents++;
2039
2040                 cur_byte = key.objectid;
2041                 num_bytes = key.offset;
2042                 ei = btrfs_item_ptr(leaf, path.slots[0],
2043                                     struct btrfs_extent_item);
2044                 BUG_ON(!(btrfs_extent_flags(leaf, ei) &
2045                          BTRFS_EXTENT_FLAG_DATA));
2046
2047                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2048                 end = ptr + btrfs_item_size_nr(leaf, path.slots[0]);
2049
2050                 ptr += sizeof(struct btrfs_extent_item);
2051
2052                 while (ptr < end) {
2053                         iref = (struct btrfs_extent_inline_ref *)ptr;
2054                         key.type = btrfs_extent_inline_ref_type(leaf, iref);
2055                         BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
2056                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2057                         ref_root = btrfs_extent_data_ref_root(leaf, dref);
2058                         extent_key.objectid =
2059                                 btrfs_extent_data_ref_objectid(leaf, dref);
2060                         extent_key.offset =
2061                                 btrfs_extent_data_ref_offset(leaf, dref);
2062                         extent_key.type = BTRFS_EXTENT_DATA_KEY;
2063                         BUG_ON(btrfs_extent_data_ref_count(leaf, dref) != 1);
2064
2065                         if (ref_root == cur_root->root_key.objectid)
2066                                 break;
2067
2068                         ptr += btrfs_extent_inline_ref_size(key.type);
2069                 }
2070
2071                 if (ptr >= end) {
2072                         path.slots[0]++;
2073                         goto next;
2074                 }
2075
2076                 ret = relocate_one_reference(trans, cur_root, cur_byte,
2077                                              num_bytes, &extent_key,
2078                                              &reloc_tree);
2079                 if (ret < 0)
2080                         goto fail;
2081
2082                 cur_byte += num_bytes;
2083                 btrfs_release_path(&path);
2084
2085                 if (trans->blocks_used >= 4096) {
2086                         ret = btrfs_commit_transaction(trans, cur_root);
2087                         BUG_ON(ret);
2088                         trans = btrfs_start_transaction(cur_root, 1);
2089                         BUG_ON(!trans);
2090                 }
2091         }
2092         btrfs_release_path(&path);
2093
2094         ret = btrfs_commit_transaction(trans, cur_root);
2095         BUG_ON(ret);
2096
2097         if (num_extents > 0 && pass++ < 16)
2098                 goto again;
2099
2100         ret = (num_extents > 0) ? -1 : 0;
2101 fail:
2102         btrfs_release_path(&path);
2103         extent_io_tree_cleanup(&reloc_tree);
2104         return ret;
2105 }
2106
2107 /*
2108  * relocate data in system chunk
2109  */
2110 static int cleanup_sys_chunk(struct btrfs_root *fs_root,
2111                              struct btrfs_root *ext2_root)
2112 {
2113         struct btrfs_block_group_cache *cache;
2114         int i, ret = 0;
2115         u64 offset = 0;
2116         u64 end_byte;
2117
2118         while(1) {
2119                 cache = btrfs_lookup_block_group(fs_root->fs_info, offset);
2120                 if (!cache)
2121                         break;
2122
2123                 end_byte = cache->key.objectid + cache->key.offset;
2124                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2125                         ret = relocate_extents_range(fs_root, ext2_root,
2126                                                      cache->key.objectid,
2127                                                      end_byte);
2128                         if (ret)
2129                                 goto fail;
2130                 }
2131                 offset = end_byte;
2132         }
2133         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2134                 offset = btrfs_sb_offset(i);
2135                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
2136
2137                 ret = relocate_extents_range(fs_root, ext2_root,
2138                                              offset, offset + BTRFS_STRIPE_LEN);
2139                 if (ret)
2140                         goto fail;
2141         }
2142         ret = 0;
2143 fail:
2144         return ret;
2145 }
2146
2147 static int fixup_chunk_mapping(struct btrfs_root *root)
2148 {
2149         struct btrfs_trans_handle *trans;
2150         struct btrfs_fs_info *info = root->fs_info;
2151         struct btrfs_root *chunk_root = info->chunk_root;
2152         struct extent_buffer *leaf;
2153         struct btrfs_key key;
2154         struct btrfs_path path;
2155         struct btrfs_chunk chunk;
2156         unsigned long ptr;
2157         u32 size;
2158         u64 type;
2159         int ret;
2160
2161         btrfs_init_path(&path);
2162
2163         trans = btrfs_start_transaction(root, 1);
2164         BUG_ON(!trans);
2165
2166         /*
2167          * recow the whole chunk tree. this will move all chunk tree blocks
2168          * into system block group.
2169          */
2170         memset(&key, 0, sizeof(key));
2171         while (1) {
2172                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2173                 if (ret < 0)
2174                         goto err;
2175
2176                 ret = btrfs_next_leaf(chunk_root, &path);
2177                 if (ret < 0)
2178                         goto err;
2179                 if (ret > 0)
2180                         break;
2181
2182                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2183                 btrfs_release_path(&path);
2184         }
2185         btrfs_release_path(&path);
2186
2187         /* fixup the system chunk array in super block */
2188         btrfs_set_super_sys_array_size(info->super_copy, 0);
2189
2190         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2191         key.offset = 0;
2192         key.type = BTRFS_CHUNK_ITEM_KEY;
2193
2194         ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 0);
2195         if (ret < 0)
2196                 goto err;
2197         BUG_ON(ret != 0);
2198         while(1) {
2199                 leaf = path.nodes[0];
2200                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2201                         ret = btrfs_next_leaf(chunk_root, &path);
2202                         if (ret < 0)
2203                                 goto err;
2204                         if (ret > 0)
2205                                 break;
2206                         leaf = path.nodes[0];
2207                 }
2208                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2209                 if (key.type != BTRFS_CHUNK_ITEM_KEY)
2210                         goto next;
2211
2212                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2213                 size = btrfs_item_size_nr(leaf, path.slots[0]);
2214                 BUG_ON(size != sizeof(chunk));
2215                 read_extent_buffer(leaf, &chunk, ptr, size);
2216                 type = btrfs_stack_chunk_type(&chunk);
2217
2218                 if (!(type & BTRFS_BLOCK_GROUP_SYSTEM))
2219                         goto next;
2220
2221                 ret = btrfs_add_system_chunk(trans, chunk_root, &key,
2222                                              &chunk, size);
2223                 if (ret)
2224                         goto err;
2225 next:
2226                 path.slots[0]++;
2227         }
2228
2229         ret = btrfs_commit_transaction(trans, root);
2230         BUG_ON(ret);
2231 err:
2232         btrfs_release_path(&path);
2233         return ret;
2234 }
2235
2236 static int do_convert(const char *devname, int datacsum, int packing, int noxattr,
2237                int copylabel, const char *fslabel, int progress)
2238 {
2239         int i, ret;
2240         int fd = -1;
2241         u32 blocksize;
2242         u64 blocks[7];
2243         u64 total_bytes;
2244         u64 super_bytenr;
2245         ext2_filsys ext2_fs;
2246         struct btrfs_root *root;
2247         struct btrfs_root *ext2_root;
2248         struct task_ctx ctx;
2249
2250         ret = open_ext2fs(devname, &ext2_fs);
2251         if (ret) {
2252                 fprintf(stderr, "unable to open the Ext2fs\n");
2253                 goto fail;
2254         }
2255         blocksize = ext2_fs->blocksize;
2256         total_bytes = (u64)ext2_fs->super->s_blocks_count * blocksize;
2257         if (blocksize < 4096) {
2258                 fprintf(stderr, "block size is too small\n");
2259                 goto fail;
2260         }
2261         if (!(ext2_fs->super->s_feature_incompat &
2262               EXT2_FEATURE_INCOMPAT_FILETYPE)) {
2263                 fprintf(stderr, "filetype feature is missing\n");
2264                 goto fail;
2265         }
2266         for (i = 0; i < 7; i++) {
2267                 ret = ext2_alloc_block(ext2_fs, 0, blocks + i);
2268                 if (ret) {
2269                         fprintf(stderr, "not enough free space\n");
2270                         goto fail;
2271                 }
2272                 blocks[i] *= blocksize;
2273         }
2274         super_bytenr = blocks[0];
2275         fd = open(devname, O_RDWR);
2276         if (fd < 0) {
2277                 fprintf(stderr, "unable to open %s\n", devname);
2278                 goto fail;
2279         }
2280         ret = make_btrfs(fd, devname, ext2_fs->super->s_volume_name,
2281                          NULL, blocks, total_bytes, blocksize, blocksize,
2282                          blocksize, blocksize, 0);
2283         if (ret) {
2284                 fprintf(stderr, "unable to create initial ctree: %s\n",
2285                         strerror(-ret));
2286                 goto fail;
2287         }
2288         /* create a system chunk that maps the whole device */
2289         ret = prepare_system_chunk(fd, super_bytenr, blocksize);
2290         if (ret) {
2291                 fprintf(stderr, "unable to update system chunk\n");
2292                 goto fail;
2293         }
2294         root = open_ctree_fd(fd, devname, super_bytenr, OPEN_CTREE_WRITES);
2295         if (!root) {
2296                 fprintf(stderr, "unable to open ctree\n");
2297                 goto fail;
2298         }
2299         ret = cache_free_extents(root, ext2_fs);
2300         if (ret) {
2301                 fprintf(stderr, "error during cache_free_extents %d\n", ret);
2302                 goto fail;
2303         }
2304         root->fs_info->extent_ops = &extent_ops;
2305         /* recover block allocation bitmap */
2306         for (i = 0; i < 7; i++) {
2307                 blocks[i] /= blocksize;
2308                 ext2_free_block(ext2_fs, blocks[i]);
2309         }
2310         ret = init_btrfs(root);
2311         if (ret) {
2312                 fprintf(stderr, "unable to setup the root tree\n");
2313                 goto fail;
2314         }
2315         printf("creating btrfs metadata.\n");
2316         ctx.max_copy_inodes = (ext2_fs->super->s_inodes_count
2317                         - ext2_fs->super->s_free_inodes_count);
2318         ctx.cur_copy_inodes = 0;
2319
2320         if (progress) {
2321                 ctx.info = task_init(print_copied_inodes, after_copied_inodes, &ctx);
2322                 task_start(ctx.info);
2323         }
2324         ret = copy_inodes(root, ext2_fs, datacsum, packing, noxattr, &ctx);
2325         if (ret) {
2326                 fprintf(stderr, "error during copy_inodes %d\n", ret);
2327                 goto fail;
2328         }
2329         if (progress) {
2330                 task_stop(ctx.info);
2331                 task_deinit(ctx.info);
2332         }
2333         printf("creating ext2fs image file.\n");
2334         ext2_root = link_subvol(root, "ext2_saved", EXT2_IMAGE_SUBVOL_OBJECTID);
2335         if (!ext2_root) {
2336                 fprintf(stderr, "unable to create subvol\n");
2337                 goto fail;
2338         }
2339         ret = create_ext2_image(ext2_root, ext2_fs, "image");
2340         if (ret) {
2341                 fprintf(stderr, "error during create_ext2_image %d\n", ret);
2342                 goto fail;
2343         }
2344         memset(root->fs_info->super_copy->label, 0, BTRFS_LABEL_SIZE);
2345         if (copylabel == 1) {
2346                 strncpy(root->fs_info->super_copy->label,
2347                                 ext2_fs->super->s_volume_name, 16);
2348                 fprintf(stderr, "copy label '%s'\n",
2349                                 root->fs_info->super_copy->label);
2350         } else if (copylabel == -1) {
2351                 strncpy(root->fs_info->super_copy->label, fslabel, BTRFS_LABEL_SIZE);
2352                 fprintf(stderr, "set label to '%s'\n", fslabel);
2353         }
2354
2355         printf("cleaning up system chunk.\n");
2356         ret = cleanup_sys_chunk(root, ext2_root);
2357         if (ret) {
2358                 fprintf(stderr, "error during cleanup_sys_chunk %d\n", ret);
2359                 goto fail;
2360         }
2361         ret = close_ctree(root);
2362         if (ret) {
2363                 fprintf(stderr, "error during close_ctree %d\n", ret);
2364                 goto fail;
2365         }
2366         close_ext2fs(ext2_fs);
2367
2368         /*
2369          * If this step succeed, we get a mountable btrfs. Otherwise
2370          * the ext2fs is left unchanged.
2371          */
2372         ret = migrate_super_block(fd, super_bytenr, blocksize);
2373         if (ret) {
2374                 fprintf(stderr, "unable to migrate super block\n");
2375                 goto fail;
2376         }
2377
2378         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2379         if (!root) {
2380                 fprintf(stderr, "unable to open ctree\n");
2381                 goto fail;
2382         }
2383         /* move chunk tree into system chunk. */
2384         ret = fixup_chunk_mapping(root);
2385         if (ret) {
2386                 fprintf(stderr, "error during fixup_chunk_tree\n");
2387                 goto fail;
2388         }
2389         ret = close_ctree(root);
2390         close(fd);
2391
2392         printf("conversion complete.\n");
2393         return 0;
2394 fail:
2395         if (fd != -1)
2396                 close(fd);
2397         fprintf(stderr, "conversion aborted.\n");
2398         return -1;
2399 }
2400
2401 static int may_rollback(struct btrfs_root *root)
2402 {
2403         struct btrfs_fs_info *info = root->fs_info;
2404         struct btrfs_multi_bio *multi = NULL;
2405         u64 bytenr;
2406         u64 length;
2407         u64 physical;
2408         u64 total_bytes;
2409         int num_stripes;
2410         int ret;
2411
2412         if (btrfs_super_num_devices(info->super_copy) != 1)
2413                 goto fail;
2414
2415         bytenr = BTRFS_SUPER_INFO_OFFSET;
2416         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
2417
2418         while (1) {
2419                 ret = btrfs_map_block(&info->mapping_tree, WRITE, bytenr,
2420                                       &length, &multi, 0, NULL);
2421                 if (ret) {
2422                         if (ret == -ENOENT) {
2423                                 /* removed block group at the tail */
2424                                 if (length == (u64)-1)
2425                                         break;
2426
2427                                 /* removed block group in the middle */
2428                                 goto next;
2429                         }
2430                         goto fail;
2431                 }
2432
2433                 num_stripes = multi->num_stripes;
2434                 physical = multi->stripes[0].physical;
2435                 kfree(multi);
2436
2437                 if (num_stripes != 1 || physical != bytenr)
2438                         goto fail;
2439 next:
2440                 bytenr += length;
2441                 if (bytenr >= total_bytes)
2442                         break;
2443         }
2444         return 0;
2445 fail:
2446         return -1;
2447 }
2448
2449 static int do_rollback(const char *devname)
2450 {
2451         int fd = -1;
2452         int ret;
2453         int i;
2454         struct btrfs_root *root;
2455         struct btrfs_root *ext2_root;
2456         struct btrfs_root *chunk_root;
2457         struct btrfs_dir_item *dir;
2458         struct btrfs_inode_item *inode;
2459         struct btrfs_file_extent_item *fi;
2460         struct btrfs_trans_handle *trans;
2461         struct extent_buffer *leaf;
2462         struct btrfs_block_group_cache *cache1;
2463         struct btrfs_block_group_cache *cache2;
2464         struct btrfs_key key;
2465         struct btrfs_path path;
2466         struct extent_io_tree io_tree;
2467         char *buf = NULL;
2468         char *name;
2469         u64 bytenr;
2470         u64 num_bytes;
2471         u64 root_dir;
2472         u64 objectid;
2473         u64 offset;
2474         u64 start;
2475         u64 end;
2476         u64 sb_bytenr;
2477         u64 first_free;
2478         u64 total_bytes;
2479         u32 sectorsize;
2480
2481         extent_io_tree_init(&io_tree);
2482
2483         fd = open(devname, O_RDWR);
2484         if (fd < 0) {
2485                 fprintf(stderr, "unable to open %s\n", devname);
2486                 goto fail;
2487         }
2488         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2489         if (!root) {
2490                 fprintf(stderr, "unable to open ctree\n");
2491                 goto fail;
2492         }
2493         ret = may_rollback(root);
2494         if (ret < 0) {
2495                 fprintf(stderr, "unable to do rollback\n");
2496                 goto fail;
2497         }
2498
2499         sectorsize = root->sectorsize;
2500         buf = malloc(sectorsize);
2501         if (!buf) {
2502                 fprintf(stderr, "unable to allocate memory\n");
2503                 goto fail;
2504         }
2505
2506         btrfs_init_path(&path);
2507
2508         key.objectid = EXT2_IMAGE_SUBVOL_OBJECTID;
2509         key.type = BTRFS_ROOT_ITEM_KEY;
2510         key.offset = (u64)-1;
2511         ext2_root = btrfs_read_fs_root(root->fs_info, &key);
2512         if (!ext2_root || IS_ERR(ext2_root)) {
2513                 fprintf(stderr, "unable to open subvol %llu\n",
2514                         (unsigned long long)key.objectid);
2515                 goto fail;
2516         }
2517
2518         name = "image";
2519         root_dir = btrfs_root_dirid(&root->root_item);
2520         dir = btrfs_lookup_dir_item(NULL, ext2_root, &path,
2521                                    root_dir, name, strlen(name), 0);
2522         if (!dir || IS_ERR(dir)) {
2523                 fprintf(stderr, "unable to find file %s\n", name);
2524                 goto fail;
2525         }
2526         leaf = path.nodes[0];
2527         btrfs_dir_item_key_to_cpu(leaf, dir, &key);
2528         btrfs_release_path(&path);
2529
2530         objectid = key.objectid;
2531
2532         ret = btrfs_lookup_inode(NULL, ext2_root, &path, &key, 0);
2533         if (ret) {
2534                 fprintf(stderr, "unable to find inode item\n");
2535                 goto fail;
2536         }
2537         leaf = path.nodes[0];
2538         inode = btrfs_item_ptr(leaf, path.slots[0], struct btrfs_inode_item);
2539         total_bytes = btrfs_inode_size(leaf, inode);
2540         btrfs_release_path(&path);
2541
2542         key.objectid = objectid;
2543         key.offset = 0;
2544         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
2545         ret = btrfs_search_slot(NULL, ext2_root, &key, &path, 0, 0);
2546         if (ret != 0) {
2547                 fprintf(stderr, "unable to find first file extent\n");
2548                 btrfs_release_path(&path);
2549                 goto fail;
2550         }
2551
2552         /* build mapping tree for the relocated blocks */
2553         for (offset = 0; offset < total_bytes; ) {
2554                 leaf = path.nodes[0];
2555                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2556                         ret = btrfs_next_leaf(root, &path);
2557                         if (ret != 0)
2558                                 break;
2559                         continue;
2560                 }
2561
2562                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2563                 if (key.objectid != objectid || key.offset != offset ||
2564                     btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2565                         break;
2566
2567                 fi = btrfs_item_ptr(leaf, path.slots[0],
2568                                     struct btrfs_file_extent_item);
2569                 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2570                         break;
2571                 if (btrfs_file_extent_compression(leaf, fi) ||
2572                     btrfs_file_extent_encryption(leaf, fi) ||
2573                     btrfs_file_extent_other_encoding(leaf, fi))
2574                         break;
2575
2576                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2577                 /* skip holes and direct mapped extents */
2578                 if (bytenr == 0 || bytenr == offset)
2579                         goto next_extent;
2580
2581                 bytenr += btrfs_file_extent_offset(leaf, fi);
2582                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
2583
2584                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2585                 cache2 =  btrfs_lookup_block_group(root->fs_info,
2586                                                    offset + num_bytes - 1);
2587                 if (!cache1 || cache1 != cache2 ||
2588                     (!(cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
2589                      !intersect_with_sb(offset, num_bytes)))
2590                         break;
2591
2592                 set_extent_bits(&io_tree, offset, offset + num_bytes - 1,
2593                                 EXTENT_LOCKED, GFP_NOFS);
2594                 set_state_private(&io_tree, offset, bytenr);
2595 next_extent:
2596                 offset += btrfs_file_extent_num_bytes(leaf, fi);
2597                 path.slots[0]++;
2598         }
2599         btrfs_release_path(&path);
2600
2601         if (offset < total_bytes) {
2602                 fprintf(stderr, "unable to build extent mapping\n");
2603                 goto fail;
2604         }
2605
2606         first_free = BTRFS_SUPER_INFO_OFFSET + 2 * sectorsize - 1;
2607         first_free &= ~((u64)sectorsize - 1);
2608         /* backup for extent #0 should exist */
2609         if(!test_range_bit(&io_tree, 0, first_free - 1, EXTENT_LOCKED, 1)) {
2610                 fprintf(stderr, "no backup for the first extent\n");
2611                 goto fail;
2612         }
2613         /* force no allocation from system block group */
2614         root->fs_info->system_allocs = -1;
2615         trans = btrfs_start_transaction(root, 1);
2616         BUG_ON(!trans);
2617         /*
2618          * recow the whole chunk tree, this will remove all chunk tree blocks
2619          * from system block group
2620          */
2621         chunk_root = root->fs_info->chunk_root;
2622         memset(&key, 0, sizeof(key));
2623         while (1) {
2624                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2625                 if (ret < 0)
2626                         break;
2627
2628                 ret = btrfs_next_leaf(chunk_root, &path);
2629                 if (ret)
2630                         break;
2631
2632                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2633                 btrfs_release_path(&path);
2634         }
2635         btrfs_release_path(&path);
2636
2637         offset = 0;
2638         num_bytes = 0;
2639         while(1) {
2640                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2641                 if (!cache1)
2642                         break;
2643
2644                 if (cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2645                         num_bytes += btrfs_block_group_used(&cache1->item);
2646
2647                 offset = cache1->key.objectid + cache1->key.offset;
2648         }
2649         /* only extent #0 left in system block group? */
2650         if (num_bytes > first_free) {
2651                 fprintf(stderr, "unable to empty system block group\n");
2652                 goto fail;
2653         }
2654         /* create a system chunk that maps the whole device */
2655         ret = prepare_system_chunk_sb(root->fs_info->super_copy);
2656         if (ret) {
2657                 fprintf(stderr, "unable to update system chunk\n");
2658                 goto fail;
2659         }
2660
2661         ret = btrfs_commit_transaction(trans, root);
2662         BUG_ON(ret);
2663
2664         ret = close_ctree(root);
2665         if (ret) {
2666                 fprintf(stderr, "error during close_ctree %d\n", ret);
2667                 goto fail;
2668         }
2669
2670         /* zero btrfs super block mirrors */
2671         memset(buf, 0, sectorsize);
2672         for (i = 1 ; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2673                 bytenr = btrfs_sb_offset(i);
2674                 if (bytenr >= total_bytes)
2675                         break;
2676                 ret = pwrite(fd, buf, sectorsize, bytenr);
2677         }
2678
2679         sb_bytenr = (u64)-1;
2680         /* copy all relocated blocks back */
2681         while(1) {
2682                 ret = find_first_extent_bit(&io_tree, 0, &start, &end,
2683                                             EXTENT_LOCKED);
2684                 if (ret)
2685                         break;
2686
2687                 ret = get_state_private(&io_tree, start, &bytenr);
2688                 BUG_ON(ret);
2689
2690                 clear_extent_bits(&io_tree, start, end, EXTENT_LOCKED,
2691                                   GFP_NOFS);
2692
2693                 while (start <= end) {
2694                         if (start == BTRFS_SUPER_INFO_OFFSET) {
2695                                 sb_bytenr = bytenr;
2696                                 goto next_sector;
2697                         }
2698                         ret = pread(fd, buf, sectorsize, bytenr);
2699                         if (ret < 0) {
2700                                 fprintf(stderr, "error during pread %d\n", ret);
2701                                 goto fail;
2702                         }
2703                         BUG_ON(ret != sectorsize);
2704                         ret = pwrite(fd, buf, sectorsize, start);
2705                         if (ret < 0) {
2706                                 fprintf(stderr, "error during pwrite %d\n", ret);
2707                                 goto fail;
2708                         }
2709                         BUG_ON(ret != sectorsize);
2710 next_sector:
2711                         start += sectorsize;
2712                         bytenr += sectorsize;
2713                 }
2714         }
2715
2716         ret = fsync(fd);
2717         if (ret) {
2718                 fprintf(stderr, "error during fsync %d\n", ret);
2719                 goto fail;
2720         }
2721         /*
2722          * finally, overwrite btrfs super block.
2723          */
2724         ret = pread(fd, buf, sectorsize, sb_bytenr);
2725         if (ret < 0) {
2726                 fprintf(stderr, "error during pread %d\n", ret);
2727                 goto fail;
2728         }
2729         BUG_ON(ret != sectorsize);
2730         ret = pwrite(fd, buf, sectorsize, BTRFS_SUPER_INFO_OFFSET);
2731         if (ret < 0) {
2732                 fprintf(stderr, "error during pwrite %d\n", ret);
2733                 goto fail;
2734         }
2735         BUG_ON(ret != sectorsize);
2736         ret = fsync(fd);
2737         if (ret) {
2738                 fprintf(stderr, "error during fsync %d\n", ret);
2739                 goto fail;
2740         }
2741
2742         close(fd);
2743         free(buf);
2744         extent_io_tree_cleanup(&io_tree);
2745         printf("rollback complete.\n");
2746         return 0;
2747
2748 fail:
2749         if (fd != -1)
2750                 close(fd);
2751         free(buf);
2752         fprintf(stderr, "rollback aborted.\n");
2753         return -1;
2754 }
2755
2756 static void print_usage(void)
2757 {
2758         printf("usage: btrfs-convert [-d] [-i] [-n] [-r] [-l label] [-L] device\n");
2759         printf("\t-d           disable data checksum\n");
2760         printf("\t-i           ignore xattrs and ACLs\n");
2761         printf("\t-n           disable packing of small files\n");
2762         printf("\t-r           roll back to ext2fs\n");
2763         printf("\t-l LABEL     set filesystem label\n");
2764         printf("\t-L           use label from converted fs\n");
2765         printf("\t-p           show converting progress\n");
2766 }
2767
2768 int main(int argc, char *argv[])
2769 {
2770         int ret;
2771         int packing = 1;
2772         int noxattr = 0;
2773         int datacsum = 1;
2774         int rollback = 0;
2775         int copylabel = 0;
2776         int usage_error = 0;
2777         int progress = 0;
2778         char *file;
2779         char *fslabel = NULL;
2780
2781         while(1) {
2782                 int c = getopt(argc, argv, "dinrl:Lp");
2783                 if (c < 0)
2784                         break;
2785                 switch(c) {
2786                         case 'd':
2787                                 datacsum = 0;
2788                                 break;
2789                         case 'i':
2790                                 noxattr = 1;
2791                                 break;
2792                         case 'n':
2793                                 packing = 0;
2794                                 break;
2795                         case 'r':
2796                                 rollback = 1;
2797                                 break;
2798                         case 'l':
2799                                 copylabel = -1;
2800                                 fslabel = strdup(optarg);
2801                                 if (strlen(fslabel) > BTRFS_LABEL_SIZE) {
2802                                         fprintf(stderr,
2803                                                 "warning: label too long, trimmed to %d bytes\n",
2804                                                 BTRFS_LABEL_SIZE);
2805                                         fslabel[BTRFS_LABEL_SIZE] = 0;
2806                                 }
2807                                 break;
2808                         case 'L':
2809                                 copylabel = 1;
2810                                 break;
2811                         case 'p':
2812                                 progress = 1;
2813                                 break;
2814                         default:
2815                                 print_usage();
2816                                 return 1;
2817                 }
2818         }
2819         argc = argc - optind;
2820         set_argv0(argv);
2821         if (check_argc_exact(argc, 1)) {
2822                 print_usage();
2823                 return 1;
2824         }
2825
2826         if (rollback && (!datacsum || noxattr || !packing)) {
2827                 fprintf(stderr,
2828                         "Usage error: -d, -i, -n options do not apply to rollback\n");
2829                 usage_error++;
2830         }
2831
2832         if (usage_error) {
2833                 print_usage();
2834                 return 1;
2835         }
2836
2837         file = argv[optind];
2838         ret = check_mounted(file);
2839         if (ret < 0) {
2840                 fprintf(stderr, "Could not check mount status: %s\n",
2841                         strerror(-ret));
2842                 return 1;
2843         } else if (ret) {
2844                 fprintf(stderr, "%s is mounted\n", file);
2845                 return 1;
2846         }
2847
2848         if (rollback) {
2849                 ret = do_rollback(file);
2850         } else {
2851                 ret = do_convert(file, datacsum, packing, noxattr, copylabel, fslabel, progress);
2852         }
2853         if (ret)
2854                 return 1;
2855         return 0;
2856 }