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 #include "kerncompat.h"
  20
  21 #include <sys/ioctl.h>
  22 #include <sys/mount.h>
  23 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <sys/types.h>
  26 #include <sys/stat.h>
  27 #include <sys/acl.h>
  28 #include <fcntl.h>
  29 #include <unistd.h>
  30 #include <uuid/uuid.h>
  31 #include <linux/limits.h>
  32
  33 #include "ctree.h"
  34 #include "disk-io.h"
  35 #include "volumes.h"
  36 #include "transaction.h"
  37 #include "crc32c.h"
  38 #include "utils.h"
  39 #include "task-utils.h"
  40 #include <ext2fs/ext2_fs.h>
  41 #include <ext2fs/ext2fs.h>
  42 #include <ext2fs/ext2_ext_attr.h>
  43
  44 #define INO_OFFSET (BTRFS_FIRST_FREE_OBJECTID - EXT2_ROOT_INO)
  45 #define EXT2_IMAGE_SUBVOL_OBJECTID BTRFS_FIRST_FREE_OBJECTID
  46
  47 struct task_ctx {
  48         uint32_t max_copy_inodes;
  49         uint32_t cur_copy_inodes;
  50         struct task_info *info;
  51 };
  52
  53 static void *print_copied_inodes(void *p)
  54 {
  55         struct task_ctx *priv = p;
  56         const char work_indicator[] = { '.', 'o', 'O', 'o' };
  57         uint32_t count = 0;
  58
  59         task_period_start(priv->info, 1000 /* 1s */);
  60         while (1) {
  61                 count++;
  62                 printf("copy inodes [%c] [%10d/%10d]\r",
  63                        work_indicator[count % 4], priv->cur_copy_inodes,
  64                        priv->max_copy_inodes);
  65                 fflush(stdout);
  66                 task_period_wait(priv->info);
  67         }
  68
  69         return NULL;
  70 }
  71
  72 static int after_copied_inodes(void *p)
  73 {
  74         struct task_ctx *priv = p;
  75
  76         printf("\n");
  77         task_period_stop(priv->info);
  78
  79         return 0;
  80 }
  81
  82 /*
  83  * Open Ext2fs in readonly mode, read block allocation bitmap and
  84  * inode bitmap into memory.
  85  */
  86 static int open_ext2fs(const char *name, ext2_filsys *ret_fs)
  87 {
  88         errcode_t ret;
  89         ext2_filsys ext2_fs;
  90         ext2_ino_t ino;
  91         ret = ext2fs_open(name, 0, 0, 0, unix_io_manager, &ext2_fs);
  92         if (ret) {
  93                 fprintf(stderr, "ext2fs_open: %s\n", error_message(ret));
  94                 goto fail;
  95         }
  96         ret = ext2fs_read_inode_bitmap(ext2_fs);
  97         if (ret) {
  98                 fprintf(stderr, "ext2fs_read_inode_bitmap: %s\n",
  99                         error_message(ret));
 100                 goto fail;
 101         }
 102         ret = ext2fs_read_block_bitmap(ext2_fs);
 103         if (ret) {
 104                 fprintf(stderr, "ext2fs_read_block_bitmap: %s\n",
 105                         error_message(ret));
 106                 goto fail;
 107         }
 108         /*
 109          * search each block group for a free inode. this set up
 110          * uninit block/inode bitmaps appropriately.
 111          */
 112         ino = 1;
 113         while (ino <= ext2_fs->super->s_inodes_count) {
 114                 ext2_ino_t foo;
 115                 ext2fs_new_inode(ext2_fs, ino, 0, NULL, &foo);
 116                 ino += EXT2_INODES_PER_GROUP(ext2_fs->super);
 117         }
 118
 119         *ret_fs = ext2_fs;
 120         return 0;
 121 fail:
 122         return -1;
 123 }
 124
 125 static int close_ext2fs(ext2_filsys fs)
 126 {
 127         ext2fs_close(fs);
 128         return 0;
 129 }
 130
 131 static int ext2_alloc_block(ext2_filsys fs, u64 goal, u64 *block_ret)
 132 {
 133         blk_t block;
 134
 135         if (!ext2fs_new_block(fs, goal, NULL, &block)) {
 136                 ext2fs_fast_mark_block_bitmap(fs->block_map, block);
 137                 *block_ret = block;
 138                 return 0;
 139         }
 140         return -ENOSPC;
 141 }
 142
 143 static int ext2_free_block(ext2_filsys fs, u64 block)
 144 {
 145         BUG_ON(block != (blk_t)block);
 146         ext2fs_fast_unmark_block_bitmap(fs->block_map, block);
 147         return 0;
 148 }
 149
 150 static int cache_free_extents(struct btrfs_root *root, ext2_filsys ext2_fs)
 151
 152 {
 153         int i, ret = 0;
 154         blk_t block;
 155         u64 bytenr;
 156         u64 blocksize = ext2_fs->blocksize;
 157
 158         block = ext2_fs->super->s_first_data_block;
 159         for (; block < ext2_fs->super->s_blocks_count; block++) {
 160                 if (ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
 161                         continue;
 162                 bytenr = block * blocksize;
 163                 ret = set_extent_dirty(&root->fs_info->free_space_cache,
 164                                        bytenr, bytenr + blocksize - 1, 0);
 165                 BUG_ON(ret);
 166         }
 167
 168         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 169                 bytenr = btrfs_sb_offset(i);
 170                 bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 171                 if (bytenr >= blocksize * ext2_fs->super->s_blocks_count)
 172                         break;
 173                 clear_extent_dirty(&root->fs_info->free_space_cache, bytenr,
 174                                    bytenr + BTRFS_STRIPE_LEN - 1, 0);
 175         }
 176
 177         clear_extent_dirty(&root->fs_info->free_space_cache,
 178                            0, BTRFS_SUPER_INFO_OFFSET - 1, 0);
 179
 180         return 0;
 181 }
 182
 183 static int custom_alloc_extent(struct btrfs_root *root, u64 num_bytes,
 184                                u64 hint_byte, struct btrfs_key *ins)
 185 {
 186         u64 start;
 187         u64 end;
 188         u64 last = hint_byte;
 189         int ret;
 190         int wrapped = 0;
 191         struct btrfs_block_group_cache *cache;
 192
 193         while(1) {
 194                 ret = find_first_extent_bit(&root->fs_info->free_space_cache,
 195                                             last, &start, &end, EXTENT_DIRTY);
 196                 if (ret) {
 197                         if (wrapped++ == 0) {
 198                                 last = 0;
 199                                 continue;
 200                         } else {
 201                                 goto fail;
 202                         }
 203                 }
 204
 205                 start = max(last, start);
 206                 last = end + 1;
 207                 if (last - start < num_bytes)
 208                         continue;
 209
 210                 last = start + num_bytes;
 211                 if (test_range_bit(&root->fs_info->pinned_extents,
 212                                    start, last - 1, EXTENT_DIRTY, 0))
 213                         continue;
 214
 215                 cache = btrfs_lookup_block_group(root->fs_info, start);
 216                 BUG_ON(!cache);
 217                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM ||
 218                     last > cache->key.objectid + cache->key.offset) {
 219                         last = cache->key.objectid + cache->key.offset;
 220                         continue;
 221                 }
 222
 223                 clear_extent_dirty(&root->fs_info->free_space_cache,
 224                                    start, start + num_bytes - 1, 0);
 225
 226                 ins->objectid = start;
 227                 ins->offset = num_bytes;
 228                 ins->type = BTRFS_EXTENT_ITEM_KEY;
 229                 return 0;
 230         }
 231 fail:
 232         fprintf(stderr, "not enough free space\n");
 233         return -ENOSPC;
 234 }
 235
 236 static int intersect_with_sb(u64 bytenr, u64 num_bytes)
 237 {
 238         int i;
 239         u64 offset;
 240
 241         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 242                 offset = btrfs_sb_offset(i);
 243                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
 244
 245                 if (bytenr < offset + BTRFS_STRIPE_LEN &&
 246                     bytenr + num_bytes > offset)
 247                         return 1;
 248         }
 249         return 0;
 250 }
 251
 252 static int custom_free_extent(struct btrfs_root *root, u64 bytenr,
 253                               u64 num_bytes)
 254 {
 255         return intersect_with_sb(bytenr, num_bytes);
 256 }
 257
 258 static struct btrfs_extent_ops extent_ops = {
 259         .alloc_extent = custom_alloc_extent,
 260         .free_extent = custom_free_extent,
 261 };
 262
 263 struct dir_iterate_data {
 264         struct btrfs_trans_handle *trans;
 265         struct btrfs_root *root;
 266         struct btrfs_inode_item *inode;
 267         u64 objectid;
 268         u64 index_cnt;
 269         u64 parent;
 270         int errcode;
 271 };
 272
 273 static u8 filetype_conversion_table[EXT2_FT_MAX] = {
 274         [EXT2_FT_UNKNOWN]       = BTRFS_FT_UNKNOWN,
 275         [EXT2_FT_REG_FILE]      = BTRFS_FT_REG_FILE,
 276         [EXT2_FT_DIR]           = BTRFS_FT_DIR,
 277         [EXT2_FT_CHRDEV]        = BTRFS_FT_CHRDEV,
 278         [EXT2_FT_BLKDEV]        = BTRFS_FT_BLKDEV,
 279         [EXT2_FT_FIFO]          = BTRFS_FT_FIFO,
 280         [EXT2_FT_SOCK]          = BTRFS_FT_SOCK,
 281         [EXT2_FT_SYMLINK]       = BTRFS_FT_SYMLINK,
 282 };
 283
 284 static int dir_iterate_proc(ext2_ino_t dir, int entry,
 285                             struct ext2_dir_entry *dirent,
 286                             int offset, int blocksize,
 287                             char *buf,void *priv_data)
 288 {
 289         int ret;
 290         int file_type;
 291         u64 objectid;
 292         u64 inode_size;
 293         char dotdot[] = "..";
 294         struct btrfs_key location;
 295         struct dir_iterate_data *idata = (struct dir_iterate_data *)priv_data;
 296         int name_len;
 297
 298         name_len = dirent->name_len & 0xFF;
 299
 300         objectid = dirent->inode + INO_OFFSET;
 301         if (!strncmp(dirent->name, dotdot, name_len)) {
 302                 if (name_len == 2) {
 303                         BUG_ON(idata->parent != 0);
 304                         idata->parent = objectid;
 305                 }
 306                 return 0;
 307         }
 308         if (dirent->inode < EXT2_GOOD_OLD_FIRST_INO)
 309                 return 0;
 310
 311         location.objectid = objectid;
 312         location.offset = 0;
 313         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
 314
 315         file_type = dirent->name_len >> 8;
 316         BUG_ON(file_type > EXT2_FT_SYMLINK);
 317         ret = btrfs_insert_dir_item(idata->trans, idata->root,
 318                                     dirent->name, name_len,
 319                                     idata->objectid, &location,
 320                                     filetype_conversion_table[file_type],
 321                                     idata->index_cnt);
 322         if (ret)
 323                 goto fail;
 324         ret = btrfs_insert_inode_ref(idata->trans, idata->root,
 325                                      dirent->name, name_len,
 326                                      objectid, idata->objectid,
 327                                      idata->index_cnt);
 328         if (ret)
 329                 goto fail;
 330         idata->index_cnt++;
 331         inode_size = btrfs_stack_inode_size(idata->inode) +
 332                      name_len * 2;
 333         btrfs_set_stack_inode_size(idata->inode, inode_size);
 334         return 0;
 335 fail:
 336         idata->errcode = ret;
 337         return BLOCK_ABORT;
 338 }
 339
 340 static int create_dir_entries(struct btrfs_trans_handle *trans,
 341                               struct btrfs_root *root, u64 objectid,
 342                               struct btrfs_inode_item *btrfs_inode,
 343                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 344 {
 345         int ret;
 346         errcode_t err;
 347         struct dir_iterate_data data = {
 348                 .trans          = trans,
 349                 .root           = root,
 350                 .inode          = btrfs_inode,
 351                 .objectid       = objectid,
 352                 .index_cnt      = 2,
 353                 .parent         = 0,
 354                 .errcode        = 0,
 355         };
 356
 357         err = ext2fs_dir_iterate2(ext2_fs, ext2_ino, 0, NULL,
 358                                   dir_iterate_proc, &data);
 359         if (err)
 360                 goto error;
 361         ret = data.errcode;
 362         if (ret == 0 && data.parent == objectid) {
 363                 ret = btrfs_insert_inode_ref(trans, root, "..", 2,
 364                                              objectid, objectid, 0);
 365         }
 366         return ret;
 367 error:
 368         fprintf(stderr, "ext2fs_dir_iterate2: %s\n", error_message(err));
 369         return -1;
 370 }
 371
 372 static int read_disk_extent(struct btrfs_root *root, u64 bytenr,
 373                             u32 num_bytes, char *buffer)
 374 {
 375         int ret;
 376         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
 377
 378         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, bytenr);
 379         if (ret != num_bytes)
 380                 goto fail;
 381         ret = 0;
 382 fail:
 383         if (ret > 0)
 384                 ret = -1;
 385         return ret;
 386 }
 387
 388 static int csum_disk_extent(struct btrfs_trans_handle *trans,
 389                             struct btrfs_root *root,
 390                             u64 disk_bytenr, u64 num_bytes)
 391 {
 392         u32 blocksize = root->sectorsize;
 393         u64 offset;
 394         char *buffer;
 395         int ret = 0;
 396
 397         buffer = malloc(blocksize);
 398         if (!buffer)
 399                 return -ENOMEM;
 400         for (offset = 0; offset < num_bytes; offset += blocksize) {
 401                 ret = read_disk_extent(root, disk_bytenr + offset,
 402                                         blocksize, buffer);
 403                 if (ret)
 404                         break;
 405                 ret = btrfs_csum_file_block(trans,
 406                                             root->fs_info->csum_root,
 407                                             disk_bytenr + num_bytes,
 408                                             disk_bytenr + offset,
 409                                             buffer, blocksize);
 410                 if (ret)
 411                         break;
 412         }
 413         free(buffer);
 414         return ret;
 415 }
 416
 417 static int record_file_blocks(struct btrfs_trans_handle *trans,
 418                               struct btrfs_root *root, u64 objectid,
 419                               struct btrfs_inode_item *inode,
 420                               u64 file_block, u64 disk_block,
 421                               u64 num_blocks, int checksum)
 422 {
 423         int ret;
 424         u64 file_pos = file_block * root->sectorsize;
 425         u64 disk_bytenr = disk_block * root->sectorsize;
 426         u64 num_bytes = num_blocks * root->sectorsize;
 427         ret = btrfs_record_file_extent(trans, root, objectid, inode, file_pos,
 428                                         disk_bytenr, num_bytes);
 429
 430         if (ret || !checksum || disk_bytenr == 0)
 431                 return ret;
 432
 433         return csum_disk_extent(trans, root, disk_bytenr, num_bytes);
 434 }
 435
 436 struct blk_iterate_data {
 437         struct btrfs_trans_handle *trans;
 438         struct btrfs_root *root;
 439         struct btrfs_inode_item *inode;
 440         u64 objectid;
 441         u64 first_block;
 442         u64 disk_block;
 443         u64 num_blocks;
 444         u64 boundary;
 445         int checksum;
 446         int errcode;
 447 };
 448
 449 static int block_iterate_proc(ext2_filsys ext2_fs,
 450                               u64 disk_block, u64 file_block,
 451                               struct blk_iterate_data *idata)
 452 {
 453         int ret;
 454         int sb_region;
 455         int do_barrier;
 456         struct btrfs_root *root = idata->root;
 457         struct btrfs_trans_handle *trans = idata->trans;
 458         struct btrfs_block_group_cache *cache;
 459         u64 bytenr = disk_block * root->sectorsize;
 460
 461         sb_region = intersect_with_sb(bytenr, root->sectorsize);
 462         do_barrier = sb_region || disk_block >= idata->boundary;
 463         if ((idata->num_blocks > 0 && do_barrier) ||
 464             (file_block > idata->first_block + idata->num_blocks) ||
 465             (disk_block != idata->disk_block + idata->num_blocks)) {
 466                 if (idata->num_blocks > 0) {
 467                         ret = record_file_blocks(trans, root, idata->objectid,
 468                                         idata->inode, idata->first_block,
 469                                         idata->disk_block, idata->num_blocks,
 470                                         idata->checksum);
 471                         if (ret)
 472                                 goto fail;
 473                         idata->first_block += idata->num_blocks;
 474                         idata->num_blocks = 0;
 475                 }
 476                 if (file_block > idata->first_block) {
 477                         ret = record_file_blocks(trans, root, idata->objectid,
 478                                         idata->inode, idata->first_block,
 479                                         0, file_block - idata->first_block,
 480                                         idata->checksum);
 481                         if (ret)
 482                                 goto fail;
 483                 }
 484
 485                 if (sb_region) {
 486                         bytenr += BTRFS_STRIPE_LEN - 1;
 487                         bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 488                 } else {
 489                         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
 490                         BUG_ON(!cache);
 491                         bytenr = cache->key.objectid + cache->key.offset;
 492                 }
 493
 494                 idata->first_block = file_block;
 495                 idata->disk_block = disk_block;
 496                 idata->boundary = bytenr / root->sectorsize;
 497         }
 498         idata->num_blocks++;
 499         return 0;
 500 fail:
 501         idata->errcode = ret;
 502         return BLOCK_ABORT;
 503 }
 504
 505 static int __block_iterate_proc(ext2_filsys fs, blk_t *blocknr,
 506                                 e2_blkcnt_t blockcnt, blk_t ref_block,
 507                                 int ref_offset, void *priv_data)
 508 {
 509         struct blk_iterate_data *idata;
 510         idata = (struct blk_iterate_data *)priv_data;
 511         return block_iterate_proc(fs, *blocknr, blockcnt, idata);
 512 }
 513
 514 /*
 515  * traverse file's data blocks, record these data blocks as file extents.
 516  */
 517 static int create_file_extents(struct btrfs_trans_handle *trans,
 518                                struct btrfs_root *root, u64 objectid,
 519                                struct btrfs_inode_item *btrfs_inode,
 520                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 521                                int datacsum, int packing)
 522 {
 523         int ret;
 524         char *buffer = NULL;
 525         errcode_t err;
 526         u32 last_block;
 527         u32 sectorsize = root->sectorsize;
 528         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 529         struct blk_iterate_data data = {
 530                 .trans          = trans,
 531                 .root           = root,
 532                 .inode          = btrfs_inode,
 533                 .objectid       = objectid,
 534                 .first_block    = 0,
 535                 .disk_block     = 0,
 536                 .num_blocks     = 0,
 537                 .boundary       = (u64)-1,
 538                 .checksum       = datacsum,
 539                 .errcode        = 0,
 540         };
 541         err = ext2fs_block_iterate2(ext2_fs, ext2_ino, BLOCK_FLAG_DATA_ONLY,
 542                                     NULL, __block_iterate_proc, &data);
 543         if (err)
 544                 goto error;
 545         ret = data.errcode;
 546         if (ret)
 547                 goto fail;
 548         if (packing && data.first_block == 0 && data.num_blocks > 0 &&
 549             inode_size <= BTRFS_MAX_INLINE_DATA_SIZE(root)) {
 550                 u64 num_bytes = data.num_blocks * sectorsize;
 551                 u64 disk_bytenr = data.disk_block * sectorsize;
 552                 u64 nbytes;
 553
 554                 buffer = malloc(num_bytes);
 555                 if (!buffer)
 556                         return -ENOMEM;
 557                 ret = read_disk_extent(root, disk_bytenr, num_bytes, buffer);
 558                 if (ret)
 559                         goto fail;
 560                 if (num_bytes > inode_size)
 561                         num_bytes = inode_size;
 562                 ret = btrfs_insert_inline_extent(trans, root, objectid,
 563                                                  0, buffer, num_bytes);
 564                 if (ret)
 565                         goto fail;
 566                 nbytes = btrfs_stack_inode_nbytes(btrfs_inode) + num_bytes;
 567                 btrfs_set_stack_inode_nbytes(btrfs_inode, nbytes);
 568         } else if (data.num_blocks > 0) {
 569                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 570                                          data.first_block, data.disk_block,
 571                                          data.num_blocks, data.checksum);
 572                 if (ret)
 573                         goto fail;
 574         }
 575         data.first_block += data.num_blocks;
 576         last_block = (inode_size + sectorsize - 1) / sectorsize;
 577         if (last_block > data.first_block) {
 578                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 579                                          data.first_block, 0, last_block -
 580                                          data.first_block, data.checksum);
 581         }
 582 fail:
 583         free(buffer);
 584         return ret;
 585 error:
 586         fprintf(stderr, "ext2fs_block_iterate2: %s\n", error_message(err));
 587         return -1;
 588 }
 589
 590 static int create_symbol_link(struct btrfs_trans_handle *trans,
 591                               struct btrfs_root *root, u64 objectid,
 592                               struct btrfs_inode_item *btrfs_inode,
 593                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 594                               struct ext2_inode *ext2_inode)
 595 {
 596         int ret;
 597         char *pathname;
 598         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 599         if (ext2fs_inode_data_blocks(ext2_fs, ext2_inode)) {
 600                 btrfs_set_stack_inode_size(btrfs_inode, inode_size + 1);
 601                 ret = create_file_extents(trans, root, objectid, btrfs_inode,
 602                                           ext2_fs, ext2_ino, 1, 1);
 603                 btrfs_set_stack_inode_size(btrfs_inode, inode_size);
 604                 return ret;
 605         }
 606
 607         pathname = (char *)&(ext2_inode->i_block[0]);
 608         BUG_ON(pathname[inode_size] != 0);
 609         ret = btrfs_insert_inline_extent(trans, root, objectid, 0,
 610                                          pathname, inode_size + 1);
 611         btrfs_set_stack_inode_nbytes(btrfs_inode, inode_size + 1);
 612         return ret;
 613 }
 614
 615 /*
 616  * Following xattr/acl related codes are based on codes in
 617  * fs/ext3/xattr.c and fs/ext3/acl.c
 618  */
 619 #define EXT2_XATTR_BHDR(ptr) ((struct ext2_ext_attr_header *)(ptr))
 620 #define EXT2_XATTR_BFIRST(ptr) \
 621         ((struct ext2_ext_attr_entry *)(EXT2_XATTR_BHDR(ptr) + 1))
 622 #define EXT2_XATTR_IHDR(inode) \
 623         ((struct ext2_ext_attr_header *) ((void *)(inode) + \
 624                 EXT2_GOOD_OLD_INODE_SIZE + (inode)->i_extra_isize))
 625 #define EXT2_XATTR_IFIRST(inode) \
 626         ((struct ext2_ext_attr_entry *) ((void *)EXT2_XATTR_IHDR(inode) + \
 627                 sizeof(EXT2_XATTR_IHDR(inode)->h_magic)))
 628
 629 static int ext2_xattr_check_names(struct ext2_ext_attr_entry *entry,
 630                                   const void *end)
 631 {
 632         struct ext2_ext_attr_entry *next;
 633
 634         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 635                 next = EXT2_EXT_ATTR_NEXT(entry);
 636                 if ((void *)next >= end)
 637                         return -EIO;
 638                 entry = next;
 639         }
 640         return 0;
 641 }
 642
 643 static int ext2_xattr_check_block(const char *buf, size_t size)
 644 {
 645         int error;
 646         struct ext2_ext_attr_header *header = EXT2_XATTR_BHDR(buf);
 647
 648         if (header->h_magic != EXT2_EXT_ATTR_MAGIC ||
 649             header->h_blocks != 1)
 650                 return -EIO;
 651         error = ext2_xattr_check_names(EXT2_XATTR_BFIRST(buf), buf + size);
 652         return error;
 653 }
 654
 655 static int ext2_xattr_check_entry(struct ext2_ext_attr_entry *entry,
 656                                   size_t size)
 657 {
 658         size_t value_size = entry->e_value_size;
 659
 660         if (entry->e_value_block != 0 || value_size > size ||
 661             entry->e_value_offs + value_size > size)
 662                 return -EIO;
 663         return 0;
 664 }
 665
 666 #define EXT2_ACL_VERSION        0x0001
 667
 668 typedef struct {
 669         __le16          e_tag;
 670         __le16          e_perm;
 671         __le32          e_id;
 672 } ext2_acl_entry;
 673
 674 typedef struct {
 675         __le16          e_tag;
 676         __le16          e_perm;
 677 } ext2_acl_entry_short;
 678
 679 typedef struct {
 680         __le32          a_version;
 681 } ext2_acl_header;
 682
 683 static inline int ext2_acl_count(size_t size)
 684 {
 685         ssize_t s;
 686         size -= sizeof(ext2_acl_header);
 687         s = size - 4 * sizeof(ext2_acl_entry_short);
 688         if (s < 0) {
 689                 if (size % sizeof(ext2_acl_entry_short))
 690                         return -1;
 691                 return size / sizeof(ext2_acl_entry_short);
 692         } else {
 693                 if (s % sizeof(ext2_acl_entry))
 694                         return -1;
 695                 return s / sizeof(ext2_acl_entry) + 4;
 696         }
 697 }
 698
 699 #define ACL_EA_VERSION          0x0002
 700
 701 typedef struct {
 702         __le16          e_tag;
 703         __le16          e_perm;
 704         __le32          e_id;
 705 } acl_ea_entry;
 706
 707 typedef struct {
 708         __le32          a_version;
 709         acl_ea_entry    a_entries[0];
 710 } acl_ea_header;
 711
 712 static inline size_t acl_ea_size(int count)
 713 {
 714         return sizeof(acl_ea_header) + count * sizeof(acl_ea_entry);
 715 }
 716
 717 static int ext2_acl_to_xattr(void *dst, const void *src,
 718                              size_t dst_size, size_t src_size)
 719 {
 720         int i, count;
 721         const void *end = src + src_size;
 722         acl_ea_header *ext_acl = (acl_ea_header *)dst;
 723         acl_ea_entry *dst_entry = ext_acl->a_entries;
 724         ext2_acl_entry *src_entry;
 725
 726         if (src_size < sizeof(ext2_acl_header))
 727                 goto fail;
 728         if (((ext2_acl_header *)src)->a_version !=
 729             cpu_to_le32(EXT2_ACL_VERSION))
 730                 goto fail;
 731         src += sizeof(ext2_acl_header);
 732         count = ext2_acl_count(src_size);
 733         if (count <= 0)
 734                 goto fail;
 735
 736         BUG_ON(dst_size < acl_ea_size(count));
 737         ext_acl->a_version = cpu_to_le32(ACL_EA_VERSION);
 738         for (i = 0; i < count; i++, dst_entry++) {
 739                 src_entry = (ext2_acl_entry *)src;
 740                 if (src + sizeof(ext2_acl_entry_short) > end)
 741                         goto fail;
 742                 dst_entry->e_tag = src_entry->e_tag;
 743                 dst_entry->e_perm = src_entry->e_perm;
 744                 switch (le16_to_cpu(src_entry->e_tag)) {
 745                 case ACL_USER_OBJ:
 746                 case ACL_GROUP_OBJ:
 747                 case ACL_MASK:
 748                 case ACL_OTHER:
 749                         src += sizeof(ext2_acl_entry_short);
 750                         dst_entry->e_id = cpu_to_le32(ACL_UNDEFINED_ID);
 751                         break;
 752                 case ACL_USER:
 753                 case ACL_GROUP:
 754                         src += sizeof(ext2_acl_entry);
 755                         if (src > end)
 756                                 goto fail;
 757                         dst_entry->e_id = src_entry->e_id;
 758                         break;
 759                 default:
 760                         goto fail;
 761                 }
 762         }
 763         if (src != end)
 764                 goto fail;
 765         return 0;
 766 fail:
 767         return -EINVAL;
 768 }
 769
 770 static char *xattr_prefix_table[] = {
 771         [1] =   "user.",
 772         [2] =   "system.posix_acl_access",
 773         [3] =   "system.posix_acl_default",
 774         [4] =   "trusted.",
 775         [6] =   "security.",
 776 };
 777
 778 static int copy_single_xattr(struct btrfs_trans_handle *trans,
 779                              struct btrfs_root *root, u64 objectid,
 780                              struct ext2_ext_attr_entry *entry,
 781                              const void *data, u32 datalen)
 782 {
 783         int ret = 0;
 784         int name_len;
 785         int name_index;
 786         void *databuf = NULL;
 787         char namebuf[XATTR_NAME_MAX + 1];
 788
 789         name_index = entry->e_name_index;
 790         if (name_index >= ARRAY_SIZE(xattr_prefix_table) ||
 791             xattr_prefix_table[name_index] == NULL)
 792                 return -EOPNOTSUPP;
 793         name_len = strlen(xattr_prefix_table[name_index]) +
 794                    entry->e_name_len;
 795         if (name_len >= sizeof(namebuf))
 796                 return -ERANGE;
 797
 798         if (name_index == 2 || name_index == 3) {
 799                 size_t bufsize = acl_ea_size(ext2_acl_count(datalen));
 800                 databuf = malloc(bufsize);
 801                 if (!databuf)
 802                        return -ENOMEM;
 803                 ret = ext2_acl_to_xattr(databuf, data, bufsize, datalen);
 804                 if (ret)
 805                         goto out;
 806                 data = databuf;
 807                 datalen = bufsize;
 808         }
 809         strncpy(namebuf, xattr_prefix_table[name_index], XATTR_NAME_MAX);
 810         strncat(namebuf, EXT2_EXT_ATTR_NAME(entry), entry->e_name_len);
 811         if (name_len + datalen > BTRFS_LEAF_DATA_SIZE(root) -
 812             sizeof(struct btrfs_item) - sizeof(struct btrfs_dir_item)) {
 813                 fprintf(stderr, "skip large xattr on inode %Lu name %.*s\n",
 814                         objectid - INO_OFFSET, name_len, namebuf);
 815                 goto out;
 816         }
 817         ret = btrfs_insert_xattr_item(trans, root, namebuf, name_len,
 818                                       data, datalen, objectid);
 819 out:
 820         free(databuf);
 821         return ret;
 822 }
 823
 824 static int copy_extended_attrs(struct btrfs_trans_handle *trans,
 825                                struct btrfs_root *root, u64 objectid,
 826                                struct btrfs_inode_item *btrfs_inode,
 827                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 828 {
 829         int ret = 0;
 830         int inline_ea = 0;
 831         errcode_t err;
 832         u32 datalen;
 833         u32 block_size = ext2_fs->blocksize;
 834         u32 inode_size = EXT2_INODE_SIZE(ext2_fs->super);
 835         struct ext2_inode_large *ext2_inode;
 836         struct ext2_ext_attr_entry *entry;
 837         void *data;
 838         char *buffer = NULL;
 839         char inode_buf[EXT2_GOOD_OLD_INODE_SIZE];
 840
 841         if (inode_size <= EXT2_GOOD_OLD_INODE_SIZE) {
 842                 ext2_inode = (struct ext2_inode_large *)inode_buf;
 843         } else {
 844                 ext2_inode = (struct ext2_inode_large *)malloc(inode_size);
 845                 if (!ext2_inode)
 846                        return -ENOMEM;
 847         }
 848         err = ext2fs_read_inode_full(ext2_fs, ext2_ino, (void *)ext2_inode,
 849                                      inode_size);
 850         if (err) {
 851                 fprintf(stderr, "ext2fs_read_inode_full: %s\n",
 852                         error_message(err));
 853                 ret = -1;
 854                 goto out;
 855         }
 856
 857         if (ext2_ino > ext2_fs->super->s_first_ino &&
 858             inode_size > EXT2_GOOD_OLD_INODE_SIZE) {
 859                 if (EXT2_GOOD_OLD_INODE_SIZE +
 860                     ext2_inode->i_extra_isize > inode_size) {
 861                         ret = -EIO;
 862                         goto out;
 863                 }
 864                 if (ext2_inode->i_extra_isize != 0 &&
 865                     EXT2_XATTR_IHDR(ext2_inode)->h_magic ==
 866                     EXT2_EXT_ATTR_MAGIC) {
 867                         inline_ea = 1;
 868                 }
 869         }
 870         if (inline_ea) {
 871                 int total;
 872                 void *end = (void *)ext2_inode + inode_size;
 873                 entry = EXT2_XATTR_IFIRST(ext2_inode);
 874                 total = end - (void *)entry;
 875                 ret = ext2_xattr_check_names(entry, end);
 876                 if (ret)
 877                         goto out;
 878                 while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 879                         ret = ext2_xattr_check_entry(entry, total);
 880                         if (ret)
 881                                 goto out;
 882                         data = (void *)EXT2_XATTR_IFIRST(ext2_inode) +
 883                                 entry->e_value_offs;
 884                         datalen = entry->e_value_size;
 885                         ret = copy_single_xattr(trans, root, objectid,
 886                                                 entry, data, datalen);
 887                         if (ret)
 888                                 goto out;
 889                         entry = EXT2_EXT_ATTR_NEXT(entry);
 890                 }
 891         }
 892
 893         if (ext2_inode->i_file_acl == 0)
 894                 goto out;
 895
 896         buffer = malloc(block_size);
 897         if (!buffer) {
 898                 ret = -ENOMEM;
 899                 goto out;
 900         }
 901         err = ext2fs_read_ext_attr(ext2_fs, ext2_inode->i_file_acl, buffer);
 902         if (err) {
 903                 fprintf(stderr, "ext2fs_read_ext_attr: %s\n",
 904                         error_message(err));
 905                 ret = -1;
 906                 goto out;
 907         }
 908         ret = ext2_xattr_check_block(buffer, block_size);
 909         if (ret)
 910                 goto out;
 911
 912         entry = EXT2_XATTR_BFIRST(buffer);
 913         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 914                 ret = ext2_xattr_check_entry(entry, block_size);
 915                 if (ret)
 916                         goto out;
 917                 data = buffer + entry->e_value_offs;
 918                 datalen = entry->e_value_size;
 919                 ret = copy_single_xattr(trans, root, objectid,
 920                                         entry, data, datalen);
 921                 if (ret)
 922                         goto out;
 923                 entry = EXT2_EXT_ATTR_NEXT(entry);
 924         }
 925 out:
 926         free(buffer);
 927         if ((void *)ext2_inode != inode_buf)
 928                 free(ext2_inode);
 929         return ret;
 930 }
 931 #define MINORBITS       20
 932 #define MKDEV(ma, mi)   (((ma) << MINORBITS) | (mi))
 933
 934 static inline dev_t old_decode_dev(u16 val)
 935 {
 936         return MKDEV((val >> 8) & 255, val & 255);
 937 }
 938
 939 static inline dev_t new_decode_dev(u32 dev)
 940 {
 941         unsigned major = (dev & 0xfff00) >> 8;
 942         unsigned minor = (dev & 0xff) | ((dev >> 12) & 0xfff00);
 943         return MKDEV(major, minor);
 944 }
 945
 946 static int copy_inode_item(struct btrfs_inode_item *dst,
 947                            struct ext2_inode *src, u32 blocksize)
 948 {
 949         btrfs_set_stack_inode_generation(dst, 1);
 950         btrfs_set_stack_inode_size(dst, src->i_size);
 951         btrfs_set_stack_inode_nbytes(dst, 0);
 952         btrfs_set_stack_inode_block_group(dst, 0);
 953         btrfs_set_stack_inode_nlink(dst, src->i_links_count);
 954         btrfs_set_stack_inode_uid(dst, src->i_uid | (src->i_uid_high << 16));
 955         btrfs_set_stack_inode_gid(dst, src->i_gid | (src->i_gid_high << 16));
 956         btrfs_set_stack_inode_mode(dst, src->i_mode);
 957         btrfs_set_stack_inode_rdev(dst, 0);
 958         btrfs_set_stack_inode_flags(dst, 0);
 959         btrfs_set_stack_timespec_sec(&dst->atime, src->i_atime);
 960         btrfs_set_stack_timespec_nsec(&dst->atime, 0);
 961         btrfs_set_stack_timespec_sec(&dst->ctime, src->i_ctime);
 962         btrfs_set_stack_timespec_nsec(&dst->ctime, 0);
 963         btrfs_set_stack_timespec_sec(&dst->mtime, src->i_mtime);
 964         btrfs_set_stack_timespec_nsec(&dst->mtime, 0);
 965         btrfs_set_stack_timespec_sec(&dst->otime, 0);
 966         btrfs_set_stack_timespec_nsec(&dst->otime, 0);
 967
 968         if (S_ISDIR(src->i_mode)) {
 969                 btrfs_set_stack_inode_size(dst, 0);
 970                 btrfs_set_stack_inode_nlink(dst, 1);
 971         }
 972         if (S_ISREG(src->i_mode)) {
 973                 btrfs_set_stack_inode_size(dst, (u64)src->i_size_high << 32 |
 974                                            (u64)src->i_size);
 975         }
 976         if (!S_ISREG(src->i_mode) && !S_ISDIR(src->i_mode) &&
 977             !S_ISLNK(src->i_mode)) {
 978                 if (src->i_block[0]) {
 979                         btrfs_set_stack_inode_rdev(dst,
 980                                 old_decode_dev(src->i_block[0]));
 981                 } else {
 982                         btrfs_set_stack_inode_rdev(dst,
 983                                 new_decode_dev(src->i_block[1]));
 984                 }
 985         }
 986         return 0;
 987 }
 988
 989 /*
 990  * copy a single inode. do all the required works, such as cloning
 991  * inode item, creating file extents and creating directory entries.
 992  */
 993 static int copy_single_inode(struct btrfs_trans_handle *trans,
 994                              struct btrfs_root *root, u64 objectid,
 995                              ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 996                              struct ext2_inode *ext2_inode,
 997                              int datacsum, int packing, int noxattr)
 998 {
 999         int ret;
1000         struct btrfs_key inode_key;
1001         struct btrfs_inode_item btrfs_inode;
1002
1003         if (ext2_inode->i_links_count == 0)
1004                 return 0;
1005
1006         copy_inode_item(&btrfs_inode, ext2_inode, ext2_fs->blocksize);
1007         if (!datacsum && S_ISREG(ext2_inode->i_mode)) {
1008                 u32 flags = btrfs_stack_inode_flags(&btrfs_inode) |
1009                             BTRFS_INODE_NODATASUM;
1010                 btrfs_set_stack_inode_flags(&btrfs_inode, flags);
1011         }
1012
1013         switch (ext2_inode->i_mode & S_IFMT) {
1014         case S_IFREG:
1015                 ret = create_file_extents(trans, root, objectid, &btrfs_inode,
1016                                         ext2_fs, ext2_ino, datacsum, packing);
1017                 break;
1018         case S_IFDIR:
1019                 ret = create_dir_entries(trans, root, objectid, &btrfs_inode,
1020                                          ext2_fs, ext2_ino);
1021                 break;
1022         case S_IFLNK:
1023                 ret = create_symbol_link(trans, root, objectid, &btrfs_inode,
1024                                          ext2_fs, ext2_ino, ext2_inode);
1025                 break;
1026         default:
1027                 ret = 0;
1028                 break;
1029         }
1030         if (ret)
1031                 return ret;
1032
1033         if (!noxattr) {
1034                 ret = copy_extended_attrs(trans, root, objectid, &btrfs_inode,
1035                                           ext2_fs, ext2_ino);
1036                 if (ret)
1037                         return ret;
1038         }
1039         inode_key.objectid = objectid;
1040         inode_key.offset = 0;
1041         btrfs_set_key_type(&inode_key, BTRFS_INODE_ITEM_KEY);
1042         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1043         return ret;
1044 }
1045
1046 static int copy_disk_extent(struct btrfs_root *root, u64 dst_bytenr,
1047                             u64 src_bytenr, u32 num_bytes)
1048 {
1049         int ret;
1050         char *buffer;
1051         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
1052
1053         buffer = malloc(num_bytes);
1054         if (!buffer)
1055                 return -ENOMEM;
1056         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, src_bytenr);
1057         if (ret != num_bytes)
1058                 goto fail;
1059         ret = pwrite(fs_devs->latest_bdev, buffer, num_bytes, dst_bytenr);
1060         if (ret != num_bytes)
1061                 goto fail;
1062         ret = 0;
1063 fail:
1064         free(buffer);
1065         if (ret > 0)
1066                 ret = -1;
1067         return ret;
1068 }
1069 /*
1070  * scan ext2's inode bitmap and copy all used inodes.
1071  */
1072 static int copy_inodes(struct btrfs_root *root, ext2_filsys ext2_fs,
1073                        int datacsum, int packing, int noxattr, struct task_ctx *p)
1074 {
1075         int ret;
1076         errcode_t err;
1077         ext2_inode_scan ext2_scan;
1078         struct ext2_inode ext2_inode;
1079         ext2_ino_t ext2_ino;
1080         u64 objectid;
1081         struct btrfs_trans_handle *trans;
1082
1083         trans = btrfs_start_transaction(root, 1);
1084         if (!trans)
1085                 return -ENOMEM;
1086         err = ext2fs_open_inode_scan(ext2_fs, 0, &ext2_scan);
1087         if (err) {
1088                 fprintf(stderr, "ext2fs_open_inode_scan: %s\n", error_message(err));
1089                 return -1;
1090         }
1091         while (!(err = ext2fs_get_next_inode(ext2_scan, &ext2_ino,
1092                                              &ext2_inode))) {
1093                 /* no more inodes */
1094                 if (ext2_ino == 0)
1095                         break;
1096                 /* skip special inode in ext2fs */
1097                 if (ext2_ino < EXT2_GOOD_OLD_FIRST_INO &&
1098                     ext2_ino != EXT2_ROOT_INO)
1099                         continue;
1100                 objectid = ext2_ino + INO_OFFSET;
1101                 ret = copy_single_inode(trans, root,
1102                                         objectid, ext2_fs, ext2_ino,
1103                                         &ext2_inode, datacsum, packing,
1104                                         noxattr);
1105                 p->cur_copy_inodes++;
1106                 if (ret)
1107                         return ret;
1108                 if (trans->blocks_used >= 4096) {
1109                         ret = btrfs_commit_transaction(trans, root);
1110                         BUG_ON(ret);
1111                         trans = btrfs_start_transaction(root, 1);
1112                         BUG_ON(!trans);
1113                 }
1114         }
1115         if (err) {
1116                 fprintf(stderr, "ext2fs_get_next_inode: %s\n", error_message(err));
1117                 return -1;
1118         }
1119         ret = btrfs_commit_transaction(trans, root);
1120         BUG_ON(ret);
1121
1122         return ret;
1123 }
1124
1125 /*
1126  * Construct a range of ext2fs image file.
1127  * scan block allocation bitmap, find all blocks used by the ext2fs
1128  * in this range and create file extents that point to these blocks.
1129  *
1130  * Note: Before calling the function, no file extent points to blocks
1131  *       in this range
1132  */
1133 static int create_image_file_range(struct btrfs_trans_handle *trans,
1134                                    struct btrfs_root *root, u64 objectid,
1135                                    struct btrfs_inode_item *inode,
1136                                    u64 start_byte, u64 end_byte,
1137                                    ext2_filsys ext2_fs)
1138 {
1139         u32 blocksize = ext2_fs->blocksize;
1140         u32 block = start_byte / blocksize;
1141         u32 last_block = (end_byte + blocksize - 1) / blocksize;
1142         int ret = 0;
1143         struct blk_iterate_data data = {
1144                 .trans          = trans,
1145                 .root           = root,
1146                 .inode          = inode,
1147                 .objectid       = objectid,
1148                 .first_block    = block,
1149                 .disk_block     = 0,
1150                 .num_blocks     = 0,
1151                 .boundary       = (u64)-1,
1152                 .checksum       = 0,
1153                 .errcode        = 0,
1154         };
1155         for (; start_byte < end_byte; block++, start_byte += blocksize) {
1156                 if (!ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
1157                         continue;
1158                 ret = block_iterate_proc(NULL, block, block, &data);
1159                 if (ret & BLOCK_ABORT) {
1160                         ret = data.errcode;
1161                         goto fail;
1162                 }
1163         }
1164         if (data.num_blocks > 0) {
1165                 ret = record_file_blocks(trans, root, objectid, inode,
1166                                          data.first_block, data.disk_block,
1167                                          data.num_blocks, 0);
1168                 if (ret)
1169                         goto fail;
1170                 data.first_block += data.num_blocks;
1171         }
1172         if (last_block > data.first_block) {
1173                 ret = record_file_blocks(trans, root, objectid, inode,
1174                                          data.first_block, 0, last_block -
1175                                          data.first_block, 0);
1176                 if (ret)
1177                         goto fail;
1178         }
1179 fail:
1180         return ret;
1181 }
1182 /*
1183  * Create the ext2fs image file.
1184  */
1185 static int create_ext2_image(struct btrfs_root *root, ext2_filsys ext2_fs,
1186                              const char *name)
1187 {
1188         int ret;
1189         struct btrfs_key key;
1190         struct btrfs_key location;
1191         struct btrfs_path path;
1192         struct btrfs_inode_item btrfs_inode;
1193         struct btrfs_inode_item *inode_item;
1194         struct extent_buffer *leaf;
1195         struct btrfs_fs_info *fs_info = root->fs_info;
1196         struct btrfs_root *extent_root = fs_info->extent_root;
1197         struct btrfs_trans_handle *trans;
1198         struct btrfs_extent_item *ei;
1199         struct btrfs_extent_inline_ref *iref;
1200         struct btrfs_extent_data_ref *dref;
1201         u64 bytenr;
1202         u64 num_bytes;
1203         u64 objectid;
1204         u64 last_byte;
1205         u64 first_free;
1206         u64 total_bytes;
1207         u32 sectorsize = root->sectorsize;
1208
1209         total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
1210         first_free =  BTRFS_SUPER_INFO_OFFSET + sectorsize * 2 - 1;
1211         first_free &= ~((u64)sectorsize - 1);
1212
1213         memset(&btrfs_inode, 0, sizeof(btrfs_inode));
1214         btrfs_set_stack_inode_generation(&btrfs_inode, 1);
1215         btrfs_set_stack_inode_size(&btrfs_inode, total_bytes);
1216         btrfs_set_stack_inode_nlink(&btrfs_inode, 1);
1217         btrfs_set_stack_inode_nbytes(&btrfs_inode, 0);
1218         btrfs_set_stack_inode_mode(&btrfs_inode, S_IFREG | 0400);
1219         btrfs_set_stack_inode_flags(&btrfs_inode, BTRFS_INODE_NODATASUM |
1220                                     BTRFS_INODE_READONLY);
1221         btrfs_init_path(&path);
1222         trans = btrfs_start_transaction(root, 1);
1223         BUG_ON(!trans);
1224
1225         objectid = btrfs_root_dirid(&root->root_item);
1226         ret = btrfs_find_free_objectid(trans, root, objectid, &objectid);
1227         if (ret)
1228                 goto fail;
1229
1230         /*
1231          * copy blocks covered by extent #0 to new positions. extent #0 is
1232          * special, we can't rely on relocate_extents_range to relocate it.
1233          */
1234         for (last_byte = 0; last_byte < first_free; last_byte += sectorsize) {
1235                 ret = custom_alloc_extent(root, sectorsize, 0, &key);
1236                 if (ret)
1237                         goto fail;
1238                 ret = copy_disk_extent(root, key.objectid, last_byte,
1239                                        sectorsize);
1240                 if (ret)
1241                         goto fail;
1242                 ret = btrfs_record_file_extent(trans, root, objectid,
1243                                                &btrfs_inode, last_byte,
1244                                                key.objectid, sectorsize);
1245                 if (ret)
1246                         goto fail;
1247         }
1248
1249         while(1) {
1250                 key.objectid = last_byte;
1251                 key.offset = 0;
1252                 btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
1253                 ret = btrfs_search_slot(trans, fs_info->extent_root,
1254                                         &key, &path, 0, 0);
1255                 if (ret < 0)
1256                         goto fail;
1257 next:
1258                 leaf = path.nodes[0];
1259                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1260                         ret = btrfs_next_leaf(extent_root, &path);
1261                         if (ret < 0)
1262                                 goto fail;
1263                         if (ret > 0)
1264                                 break;
1265                         leaf = path.nodes[0];
1266                 }
1267                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1268                 if (last_byte > key.objectid ||
1269                     key.type != BTRFS_EXTENT_ITEM_KEY) {
1270                         path.slots[0]++;
1271                         goto next;
1272                 }
1273
1274                 bytenr = key.objectid;
1275                 num_bytes = key.offset;
1276                 ei = btrfs_item_ptr(leaf, path.slots[0],
1277                                     struct btrfs_extent_item);
1278                 if (!(btrfs_extent_flags(leaf, ei) & BTRFS_EXTENT_FLAG_DATA)) {
1279                         path.slots[0]++;
1280                         goto next;
1281                 }
1282
1283                 BUG_ON(btrfs_item_size_nr(leaf, path.slots[0]) != sizeof(*ei) +
1284                        btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY));
1285
1286                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
1287                 key.type = btrfs_extent_inline_ref_type(leaf, iref);
1288                 BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
1289                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1290                 if (btrfs_extent_data_ref_root(leaf, dref) !=
1291                     BTRFS_FS_TREE_OBJECTID) {
1292                         path.slots[0]++;
1293                         goto next;
1294                 }
1295
1296                 if (bytenr > last_byte) {
1297                         ret = create_image_file_range(trans, root, objectid,
1298                                                       &btrfs_inode, last_byte,
1299                                                       bytenr, ext2_fs);
1300                         if (ret)
1301                                 goto fail;
1302                 }
1303                 ret = btrfs_record_file_extent(trans, root, objectid,
1304                                                &btrfs_inode, bytenr, bytenr,
1305                                                num_bytes);
1306                 if (ret)
1307                         goto fail;
1308                 last_byte = bytenr + num_bytes;
1309                 btrfs_release_path(&path);
1310
1311                 if (trans->blocks_used >= 4096) {
1312                         ret = btrfs_commit_transaction(trans, root);
1313                         BUG_ON(ret);
1314                         trans = btrfs_start_transaction(root, 1);
1315                         BUG_ON(!trans);
1316                 }
1317         }
1318         btrfs_release_path(&path);
1319         if (total_bytes > last_byte) {
1320                 ret = create_image_file_range(trans, root, objectid,
1321                                               &btrfs_inode, last_byte,
1322                                               total_bytes, ext2_fs);
1323                 if (ret)
1324                         goto fail;
1325         }
1326
1327         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1328         if (ret)
1329                 goto fail;
1330
1331         location.objectid = objectid;
1332         location.offset = 0;
1333         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1334         ret = btrfs_insert_dir_item(trans, root, name, strlen(name),
1335                                     btrfs_root_dirid(&root->root_item),
1336                                     &location, EXT2_FT_REG_FILE, objectid);
1337         if (ret)
1338                 goto fail;
1339         ret = btrfs_insert_inode_ref(trans, root, name, strlen(name),
1340                                      objectid,
1341                                      btrfs_root_dirid(&root->root_item),
1342                                      objectid);
1343         if (ret)
1344                 goto fail;
1345         location.objectid = btrfs_root_dirid(&root->root_item);
1346         location.offset = 0;
1347         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1348         ret = btrfs_lookup_inode(trans, root, &path, &location, 1);
1349         if (ret)
1350                 goto fail;
1351         leaf = path.nodes[0];
1352         inode_item = btrfs_item_ptr(leaf, path.slots[0],
1353                                     struct btrfs_inode_item);
1354         btrfs_set_inode_size(leaf, inode_item, strlen(name) * 2 +
1355                              btrfs_inode_size(leaf, inode_item));
1356         btrfs_mark_buffer_dirty(leaf);
1357         btrfs_release_path(&path);
1358         ret = btrfs_commit_transaction(trans, root);
1359         BUG_ON(ret);
1360 fail:
1361         btrfs_release_path(&path);
1362         return ret;
1363 }
1364
1365 static struct btrfs_root * link_subvol(struct btrfs_root *root,
1366                 const char *base, u64 root_objectid)
1367 {
1368         struct btrfs_trans_handle *trans;
1369         struct btrfs_fs_info *fs_info = root->fs_info;
1370         struct btrfs_root *tree_root = fs_info->tree_root;
1371         struct btrfs_root *new_root = NULL;
1372         struct btrfs_path *path;
1373         struct btrfs_inode_item *inode_item;
1374         struct extent_buffer *leaf;
1375         struct btrfs_key key;
1376         u64 dirid = btrfs_root_dirid(&root->root_item);
1377         u64 index = 2;
1378         char buf[BTRFS_NAME_LEN + 1]; /* for snprintf null */
1379         int len;
1380         int i;
1381         int ret;
1382
1383         len = strlen(base);
1384         if (len < 1 || len > BTRFS_NAME_LEN)
1385                 return NULL;
1386
1387         path = btrfs_alloc_path();
1388         BUG_ON(!path);
1389
1390         key.objectid = dirid;
1391         key.type = BTRFS_DIR_INDEX_KEY;
1392         key.offset = (u64)-1;
1393
1394         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1395         BUG_ON(ret <= 0);
1396
1397         if (path->slots[0] > 0) {
1398                 path->slots[0]--;
1399                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1400                 if (key.objectid == dirid && key.type == BTRFS_DIR_INDEX_KEY)
1401                         index = key.offset + 1;
1402         }
1403         btrfs_release_path(path);
1404
1405         trans = btrfs_start_transaction(root, 1);
1406         BUG_ON(!trans);
1407
1408         key.objectid = dirid;
1409         key.offset = 0;
1410         key.type =  BTRFS_INODE_ITEM_KEY;
1411
1412         ret = btrfs_lookup_inode(trans, root, path, &key, 1);
1413         BUG_ON(ret);
1414         leaf = path->nodes[0];
1415         inode_item = btrfs_item_ptr(leaf, path->slots[0],
1416                                     struct btrfs_inode_item);
1417
1418         key.objectid = root_objectid;
1419         key.offset = (u64)-1;
1420         key.type = BTRFS_ROOT_ITEM_KEY;
1421
1422         memcpy(buf, base, len);
1423         for (i = 0; i < 1024; i++) {
1424                 ret = btrfs_insert_dir_item(trans, root, buf, len,
1425                                             dirid, &key, BTRFS_FT_DIR, index);
1426                 if (ret != -EEXIST)
1427                         break;
1428                 len = snprintf(buf, ARRAY_SIZE(buf), "%s%d", base, i);
1429                 if (len < 1 || len > BTRFS_NAME_LEN) {
1430                         ret = -EINVAL;
1431                         break;
1432                 }
1433         }
1434         if (ret)
1435                 goto fail;
1436
1437         btrfs_set_inode_size(leaf, inode_item, len * 2 +
1438                              btrfs_inode_size(leaf, inode_item));
1439         btrfs_mark_buffer_dirty(leaf);
1440         btrfs_release_path(path);
1441
1442         /* add the backref first */
1443         ret = btrfs_add_root_ref(trans, tree_root, root_objectid,
1444                                  BTRFS_ROOT_BACKREF_KEY,
1445                                  root->root_key.objectid,
1446                                  dirid, index, buf, len);
1447         BUG_ON(ret);
1448
1449         /* now add the forward ref */
1450         ret = btrfs_add_root_ref(trans, tree_root, root->root_key.objectid,
1451                                  BTRFS_ROOT_REF_KEY, root_objectid,
1452                                  dirid, index, buf, len);
1453
1454         ret = btrfs_commit_transaction(trans, root);
1455         BUG_ON(ret);
1456
1457         new_root = btrfs_read_fs_root(fs_info, &key);
1458         if (IS_ERR(new_root))
1459                 new_root = NULL;
1460 fail:
1461         btrfs_free_path(path);
1462         return new_root;
1463 }
1464
1465 static int create_chunk_mapping(struct btrfs_trans_handle *trans,
1466                                 struct btrfs_root *root)
1467 {
1468         struct btrfs_fs_info *info = root->fs_info;
1469         struct btrfs_root *chunk_root = info->chunk_root;
1470         struct btrfs_root *extent_root = info->extent_root;
1471         struct btrfs_device *device;
1472         struct btrfs_block_group_cache *cache;
1473         struct btrfs_dev_extent *extent;
1474         struct extent_buffer *leaf;
1475         struct btrfs_chunk chunk;
1476         struct btrfs_key key;
1477         struct btrfs_path path;
1478         u64 cur_start;
1479         u64 total_bytes;
1480         u64 chunk_objectid;
1481         int ret;
1482
1483         btrfs_init_path(&path);
1484
1485         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
1486         chunk_objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
1487
1488         BUG_ON(list_empty(&info->fs_devices->devices));
1489         device = list_entry(info->fs_devices->devices.next,
1490                             struct btrfs_device, dev_list);
1491         BUG_ON(device->devid != info->fs_devices->latest_devid);
1492
1493         /* delete device extent created by make_btrfs */
1494         key.objectid = device->devid;
1495         key.offset = 0;
1496         key.type = BTRFS_DEV_EXTENT_KEY;
1497         ret = btrfs_search_slot(trans, device->dev_root, &key, &path, -1, 1);
1498         if (ret < 0)
1499                 goto err;
1500
1501         BUG_ON(ret > 0);
1502         ret = btrfs_del_item(trans, device->dev_root, &path);
1503         if (ret)
1504                 goto err;
1505         btrfs_release_path(&path);
1506
1507         /* delete chunk item created by make_btrfs */
1508         key.objectid = chunk_objectid;
1509         key.offset = 0;
1510         key.type = BTRFS_CHUNK_ITEM_KEY;
1511         ret = btrfs_search_slot(trans, chunk_root, &key, &path, -1, 1);
1512         if (ret < 0)
1513                 goto err;
1514
1515         BUG_ON(ret > 0);
1516         ret = btrfs_del_item(trans, chunk_root, &path);
1517         if (ret)
1518                 goto err;
1519         btrfs_release_path(&path);
1520
1521         /* for each block group, create device extent and chunk item */
1522         cur_start = 0;
1523         while (cur_start < total_bytes) {
1524                 cache = btrfs_lookup_block_group(root->fs_info, cur_start);
1525                 BUG_ON(!cache);
1526
1527                 /* insert device extent */
1528                 key.objectid = device->devid;
1529                 key.offset = cache->key.objectid;
1530                 key.type = BTRFS_DEV_EXTENT_KEY;
1531                 ret = btrfs_insert_empty_item(trans, device->dev_root, &path,
1532                                               &key, sizeof(*extent));
1533                 if (ret)
1534                         goto err;
1535
1536                 leaf = path.nodes[0];
1537                 extent = btrfs_item_ptr(leaf, path.slots[0],
1538                                         struct btrfs_dev_extent);
1539
1540                 btrfs_set_dev_extent_chunk_tree(leaf, extent,
1541                                                 chunk_root->root_key.objectid);
1542                 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
1543                                                     chunk_objectid);
1544                 btrfs_set_dev_extent_chunk_offset(leaf, extent,
1545                                                   cache->key.objectid);
1546                 btrfs_set_dev_extent_length(leaf, extent, cache->key.offset);
1547                 write_extent_buffer(leaf, root->fs_info->chunk_tree_uuid,
1548                     (unsigned long)btrfs_dev_extent_chunk_tree_uuid(extent),
1549                     BTRFS_UUID_SIZE);
1550                 btrfs_mark_buffer_dirty(leaf);
1551                 btrfs_release_path(&path);
1552
1553                 /* insert chunk item */
1554                 btrfs_set_stack_chunk_length(&chunk, cache->key.offset);
1555                 btrfs_set_stack_chunk_owner(&chunk,
1556                                             extent_root->root_key.objectid);
1557                 btrfs_set_stack_chunk_stripe_len(&chunk, BTRFS_STRIPE_LEN);
1558                 btrfs_set_stack_chunk_type(&chunk, cache->flags);
1559                 btrfs_set_stack_chunk_io_align(&chunk, device->io_align);
1560                 btrfs_set_stack_chunk_io_width(&chunk, device->io_width);
1561                 btrfs_set_stack_chunk_sector_size(&chunk, device->sector_size);
1562                 btrfs_set_stack_chunk_num_stripes(&chunk, 1);
1563                 btrfs_set_stack_chunk_sub_stripes(&chunk, 0);
1564                 btrfs_set_stack_stripe_devid(&chunk.stripe, device->devid);
1565                 btrfs_set_stack_stripe_offset(&chunk.stripe,
1566                                               cache->key.objectid);
1567                 memcpy(&chunk.stripe.dev_uuid, device->uuid, BTRFS_UUID_SIZE);
1568
1569                 key.objectid = chunk_objectid;
1570                 key.offset = cache->key.objectid;
1571                 key.type = BTRFS_CHUNK_ITEM_KEY;
1572
1573                 ret = btrfs_insert_item(trans, chunk_root, &key, &chunk,
1574                                         btrfs_chunk_item_size(1));
1575                 if (ret)
1576                         goto err;
1577
1578                 cur_start = cache->key.objectid + cache->key.offset;
1579         }
1580
1581         device->bytes_used = total_bytes;
1582         ret = btrfs_update_device(trans, device);
1583 err:
1584         btrfs_release_path(&path);
1585         return ret;
1586 }
1587
1588 static int create_subvol(struct btrfs_trans_handle *trans,
1589                          struct btrfs_root *root, u64 root_objectid)
1590 {
1591         struct extent_buffer *tmp;
1592         struct btrfs_root *new_root;
1593         struct btrfs_key key;
1594         struct btrfs_root_item root_item;
1595         int ret;
1596
1597         ret = btrfs_copy_root(trans, root, root->node, &tmp,
1598                               root_objectid);
1599         BUG_ON(ret);
1600
1601         memcpy(&root_item, &root->root_item, sizeof(root_item));
1602         btrfs_set_root_bytenr(&root_item, tmp->start);
1603         btrfs_set_root_level(&root_item, btrfs_header_level(tmp));
1604         btrfs_set_root_generation(&root_item, trans->transid);
1605         free_extent_buffer(tmp);
1606
1607         key.objectid = root_objectid;
1608         key.type = BTRFS_ROOT_ITEM_KEY;
1609         key.offset = trans->transid;
1610         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1611                                 &key, &root_item);
1612
1613         key.offset = (u64)-1;
1614         new_root = btrfs_read_fs_root(root->fs_info, &key);
1615         BUG_ON(!new_root || IS_ERR(new_root));
1616
1617         ret = btrfs_make_root_dir(trans, new_root, BTRFS_FIRST_FREE_OBJECTID);
1618         BUG_ON(ret);
1619
1620         return 0;
1621 }
1622
1623 static int init_btrfs(struct btrfs_root *root)
1624 {
1625         int ret;
1626         struct btrfs_key location;
1627         struct btrfs_trans_handle *trans;
1628         struct btrfs_fs_info *fs_info = root->fs_info;
1629         struct extent_buffer *tmp;
1630
1631         trans = btrfs_start_transaction(root, 1);
1632         BUG_ON(!trans);
1633         ret = btrfs_make_block_groups(trans, root);
1634         if (ret)
1635                 goto err;
1636         ret = btrfs_fix_block_accounting(trans, root);
1637         if (ret)
1638                 goto err;
1639         ret = create_chunk_mapping(trans, root);
1640         if (ret)
1641                 goto err;
1642         ret = btrfs_make_root_dir(trans, fs_info->tree_root,
1643                                   BTRFS_ROOT_TREE_DIR_OBJECTID);
1644         if (ret)
1645                 goto err;
1646         memcpy(&location, &root->root_key, sizeof(location));
1647         location.offset = (u64)-1;
1648         ret = btrfs_insert_dir_item(trans, fs_info->tree_root, "default", 7,
1649                                 btrfs_super_root_dir(fs_info->super_copy),
1650                                 &location, BTRFS_FT_DIR, 0);
1651         if (ret)
1652                 goto err;
1653         ret = btrfs_insert_inode_ref(trans, fs_info->tree_root, "default", 7,
1654                                 location.objectid,
1655                                 btrfs_super_root_dir(fs_info->super_copy), 0);
1656         if (ret)
1657                 goto err;
1658         btrfs_set_root_dirid(&fs_info->fs_root->root_item,
1659                              BTRFS_FIRST_FREE_OBJECTID);
1660
1661         /* subvol for ext2 image file */
1662         ret = create_subvol(trans, root, EXT2_IMAGE_SUBVOL_OBJECTID);
1663         BUG_ON(ret);
1664         /* subvol for data relocation */
1665         ret = create_subvol(trans, root, BTRFS_DATA_RELOC_TREE_OBJECTID);
1666         BUG_ON(ret);
1667
1668         extent_buffer_get(fs_info->csum_root->node);
1669         ret = __btrfs_cow_block(trans, fs_info->csum_root,
1670                                 fs_info->csum_root->node, NULL, 0, &tmp, 0, 0);
1671         BUG_ON(ret);
1672         free_extent_buffer(tmp);
1673
1674         ret = btrfs_commit_transaction(trans, root);
1675         BUG_ON(ret);
1676 err:
1677         return ret;
1678 }
1679
1680 /*
1681  * Migrate super block to its default position and zero 0 ~ 16k
1682  */
1683 static int migrate_super_block(int fd, u64 old_bytenr, u32 sectorsize)
1684 {
1685         int ret;
1686         struct extent_buffer *buf;
1687         struct btrfs_super_block *super;
1688         u32 len;
1689         u32 bytenr;
1690
1691         BUG_ON(sectorsize < sizeof(*super));
1692         buf = malloc(sizeof(*buf) + sectorsize);
1693         if (!buf)
1694                 return -ENOMEM;
1695
1696         buf->len = sectorsize;
1697         ret = pread(fd, buf->data, sectorsize, old_bytenr);
1698         if (ret != sectorsize)
1699                 goto fail;
1700
1701         super = (struct btrfs_super_block *)buf->data;
1702         BUG_ON(btrfs_super_bytenr(super) != old_bytenr);
1703         btrfs_set_super_bytenr(super, BTRFS_SUPER_INFO_OFFSET);
1704
1705         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1706         ret = pwrite(fd, buf->data, sectorsize, BTRFS_SUPER_INFO_OFFSET);
1707         if (ret != sectorsize)
1708                 goto fail;
1709
1710         ret = fsync(fd);
1711         if (ret)
1712                 goto fail;
1713
1714         memset(buf->data, 0, sectorsize);
1715         for (bytenr = 0; bytenr < BTRFS_SUPER_INFO_OFFSET; ) {
1716                 len = BTRFS_SUPER_INFO_OFFSET - bytenr;
1717                 if (len > sectorsize)
1718                         len = sectorsize;
1719                 ret = pwrite(fd, buf->data, len, bytenr);
1720                 if (ret != len) {
1721                         fprintf(stderr, "unable to zero fill device\n");
1722                         break;
1723                 }
1724                 bytenr += len;
1725         }
1726         ret = 0;
1727         fsync(fd);
1728 fail:
1729         free(buf);
1730         if (ret > 0)
1731                 ret = -1;
1732         return ret;
1733 }
1734
1735 static int prepare_system_chunk_sb(struct btrfs_super_block *super)
1736 {
1737         struct btrfs_chunk *chunk;
1738         struct btrfs_disk_key *key;
1739         u32 sectorsize = btrfs_super_sectorsize(super);
1740
1741         key = (struct btrfs_disk_key *)(super->sys_chunk_array);
1742         chunk = (struct btrfs_chunk *)(super->sys_chunk_array +
1743                                        sizeof(struct btrfs_disk_key));
1744
1745         btrfs_set_disk_key_objectid(key, BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1746         btrfs_set_disk_key_type(key, BTRFS_CHUNK_ITEM_KEY);
1747         btrfs_set_disk_key_offset(key, 0);
1748
1749         btrfs_set_stack_chunk_length(chunk, btrfs_super_total_bytes(super));
1750         btrfs_set_stack_chunk_owner(chunk, BTRFS_EXTENT_TREE_OBJECTID);
1751         btrfs_set_stack_chunk_stripe_len(chunk, BTRFS_STRIPE_LEN);
1752         btrfs_set_stack_chunk_type(chunk, BTRFS_BLOCK_GROUP_SYSTEM);
1753         btrfs_set_stack_chunk_io_align(chunk, sectorsize);
1754         btrfs_set_stack_chunk_io_width(chunk, sectorsize);
1755         btrfs_set_stack_chunk_sector_size(chunk, sectorsize);
1756         btrfs_set_stack_chunk_num_stripes(chunk, 1);
1757         btrfs_set_stack_chunk_sub_stripes(chunk, 0);
1758         chunk->stripe.devid = super->dev_item.devid;
1759         btrfs_set_stack_stripe_offset(&chunk->stripe, 0);
1760         memcpy(chunk->stripe.dev_uuid, super->dev_item.uuid, BTRFS_UUID_SIZE);
1761         btrfs_set_super_sys_array_size(super, sizeof(*key) + sizeof(*chunk));
1762         return 0;
1763 }
1764
1765 static int prepare_system_chunk(int fd, u64 sb_bytenr, u32 sectorsize)
1766 {
1767         int ret;
1768         struct extent_buffer *buf;
1769         struct btrfs_super_block *super;
1770
1771         BUG_ON(sectorsize < sizeof(*super));
1772         buf = malloc(sizeof(*buf) + sectorsize);
1773         if (!buf)
1774                 return -ENOMEM;
1775
1776         buf->len = sectorsize;
1777         ret = pread(fd, buf->data, sectorsize, sb_bytenr);
1778         if (ret != sectorsize)
1779                 goto fail;
1780
1781         super = (struct btrfs_super_block *)buf->data;
1782         BUG_ON(btrfs_super_bytenr(super) != sb_bytenr);
1783         BUG_ON(btrfs_super_num_devices(super) != 1);
1784
1785         ret = prepare_system_chunk_sb(super);
1786         if (ret)
1787                 goto fail;
1788
1789         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1790         ret = pwrite(fd, buf->data, sectorsize, sb_bytenr);
1791         if (ret != sectorsize)
1792                 goto fail;
1793
1794         ret = 0;
1795 fail:
1796         free(buf);
1797         if (ret > 0)
1798                 ret = -1;
1799         return ret;
1800 }
1801
1802 static int relocate_one_reference(struct btrfs_trans_handle *trans,
1803                                   struct btrfs_root *root,
1804                                   u64 extent_start, u64 extent_size,
1805                                   struct btrfs_key *extent_key,
1806                                   struct extent_io_tree *reloc_tree)
1807 {
1808         struct extent_buffer *leaf;
1809         struct btrfs_file_extent_item *fi;
1810         struct btrfs_key key;
1811         struct btrfs_path path;
1812         struct btrfs_inode_item inode;
1813         struct blk_iterate_data data;
1814         u64 bytenr;
1815         u64 num_bytes;
1816         u64 cur_offset;
1817         u64 new_pos;
1818         u64 nbytes;
1819         u64 sector_end;
1820         u32 sectorsize = root->sectorsize;
1821         unsigned long ptr;
1822         int datacsum;
1823         int fd;
1824         int ret;
1825
1826         btrfs_init_path(&path);
1827         ret = btrfs_search_slot(trans, root, extent_key, &path, -1, 1);
1828         if (ret)
1829                 goto fail;
1830
1831         leaf = path.nodes[0];
1832         fi = btrfs_item_ptr(leaf, path.slots[0],
1833                             struct btrfs_file_extent_item);
1834         BUG_ON(btrfs_file_extent_offset(leaf, fi) > 0);
1835         if (extent_start != btrfs_file_extent_disk_bytenr(leaf, fi) ||
1836             extent_size != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1837                 ret = 1;
1838                 goto fail;
1839         }
1840
1841         bytenr = extent_start + btrfs_file_extent_offset(leaf, fi);
1842         num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
1843
1844         ret = btrfs_del_item(trans, root, &path);
1845         if (ret)
1846                 goto fail;
1847
1848         ret = btrfs_free_extent(trans, root, extent_start, extent_size, 0,
1849                                 root->root_key.objectid,
1850                                 extent_key->objectid, extent_key->offset);
1851         if (ret)
1852                 goto fail;
1853
1854         btrfs_release_path(&path);
1855
1856         key.objectid = extent_key->objectid;
1857         key.offset = 0;
1858         key.type =  BTRFS_INODE_ITEM_KEY;
1859         ret = btrfs_lookup_inode(trans, root, &path, &key, 0);
1860         if (ret)
1861                 goto fail;
1862
1863         leaf = path.nodes[0];
1864         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1865         read_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1866         btrfs_release_path(&path);
1867
1868         BUG_ON(num_bytes & (sectorsize - 1));
1869         nbytes = btrfs_stack_inode_nbytes(&inode) - num_bytes;
1870         btrfs_set_stack_inode_nbytes(&inode, nbytes);
1871         datacsum = !(btrfs_stack_inode_flags(&inode) & BTRFS_INODE_NODATASUM);
1872
1873         data = (struct blk_iterate_data) {
1874                 .trans          = trans,
1875                 .root           = root,
1876                 .inode          = &inode,
1877                 .objectid       = extent_key->objectid,
1878                 .first_block    = extent_key->offset / sectorsize,
1879                 .disk_block     = 0,
1880                 .num_blocks     = 0,
1881                 .boundary       = (u64)-1,
1882                 .checksum       = datacsum,
1883                 .errcode        = 0,
1884         };
1885
1886         cur_offset = extent_key->offset;
1887         while (num_bytes > 0) {
1888                 sector_end = bytenr + sectorsize - 1;
1889                 if (test_range_bit(reloc_tree, bytenr, sector_end,
1890                                    EXTENT_LOCKED, 1)) {
1891                         ret = get_state_private(reloc_tree, bytenr, &new_pos);
1892                         BUG_ON(ret);
1893                 } else {
1894                         ret = custom_alloc_extent(root, sectorsize, 0, &key);
1895                         if (ret)
1896                                 goto fail;
1897                         new_pos = key.objectid;
1898
1899                         if (cur_offset == extent_key->offset) {
1900                                 fd = root->fs_info->fs_devices->latest_bdev;
1901                                 readahead(fd, bytenr, num_bytes);
1902                         }
1903                         ret = copy_disk_extent(root, new_pos, bytenr,
1904                                                sectorsize);
1905                         if (ret)
1906                                 goto fail;
1907                         ret = set_extent_bits(reloc_tree, bytenr, sector_end,
1908                                               EXTENT_LOCKED, GFP_NOFS);
1909                         BUG_ON(ret);
1910                         ret = set_state_private(reloc_tree, bytenr, new_pos);
1911                         BUG_ON(ret);
1912                 }
1913
1914                 ret = block_iterate_proc(NULL, new_pos / sectorsize,
1915                                          cur_offset / sectorsize, &data);
1916                 if (ret & BLOCK_ABORT) {
1917                         ret = data.errcode;
1918                         goto fail;
1919                 }
1920
1921                 cur_offset += sectorsize;
1922                 bytenr += sectorsize;
1923                 num_bytes -= sectorsize;
1924         }
1925
1926         if (data.num_blocks > 0) {
1927                 ret = record_file_blocks(trans, root,
1928                                          extent_key->objectid, &inode,
1929                                          data.first_block, data.disk_block,
1930                                          data.num_blocks, datacsum);
1931                 if (ret)
1932                         goto fail;
1933         }
1934
1935         key.objectid = extent_key->objectid;
1936         key.offset = 0;
1937         key.type =  BTRFS_INODE_ITEM_KEY;
1938         ret = btrfs_lookup_inode(trans, root, &path, &key, 1);
1939         if (ret)
1940                 goto fail;
1941
1942         leaf = path.nodes[0];
1943         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1944         write_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1945         btrfs_mark_buffer_dirty(leaf);
1946         btrfs_release_path(&path);
1947
1948 fail:
1949         btrfs_release_path(&path);
1950         return ret;
1951 }
1952
1953 static int relocate_extents_range(struct btrfs_root *fs_root,
1954                                   struct btrfs_root *ext2_root,
1955                                   u64 start_byte, u64 end_byte)
1956 {
1957         struct btrfs_fs_info *info = fs_root->fs_info;
1958         struct btrfs_root *extent_root = info->extent_root;
1959         struct btrfs_root *cur_root = NULL;
1960         struct btrfs_trans_handle *trans;
1961         struct btrfs_extent_data_ref *dref;
1962         struct btrfs_extent_inline_ref *iref;
1963         struct btrfs_extent_item *ei;
1964         struct extent_buffer *leaf;
1965         struct btrfs_key key;
1966         struct btrfs_key extent_key;
1967         struct btrfs_path path;
1968         struct extent_io_tree reloc_tree;
1969         unsigned long ptr;
1970         unsigned long end;
1971         u64 cur_byte;
1972         u64 num_bytes;
1973         u64 ref_root;
1974         u64 num_extents;
1975         int pass = 0;
1976         int ret;
1977
1978         btrfs_init_path(&path);
1979         extent_io_tree_init(&reloc_tree);
1980
1981         key.objectid = start_byte;
1982         key.offset = 0;
1983         key.type = BTRFS_EXTENT_ITEM_KEY;
1984         ret = btrfs_search_slot(NULL, extent_root, &key, &path, 0, 0);
1985         if (ret < 0)
1986                 goto fail;
1987         if (ret > 0) {
1988                 ret = btrfs_previous_item(extent_root, &path, 0,
1989                                           BTRFS_EXTENT_ITEM_KEY);
1990                 if (ret < 0)
1991                         goto fail;
1992                 if (ret == 0) {
1993                         leaf = path.nodes[0];
1994                         btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1995                         if (key.objectid + key.offset > start_byte)
1996                                 start_byte = key.objectid;
1997                 }
1998         }
1999         btrfs_release_path(&path);
2000 again:
2001         cur_root = (pass % 2 == 0) ? ext2_root : fs_root;
2002         num_extents = 0;
2003
2004         trans = btrfs_start_transaction(cur_root, 1);
2005         BUG_ON(!trans);
2006
2007         cur_byte = start_byte;
2008         while (1) {
2009                 key.objectid = cur_byte;
2010                 key.offset = 0;
2011                 key.type = BTRFS_EXTENT_ITEM_KEY;
2012                 ret = btrfs_search_slot(trans, extent_root,
2013                                         &key, &path, 0, 0);
2014                 if (ret < 0)
2015                         goto fail;
2016 next:
2017                 leaf = path.nodes[0];
2018                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2019                         ret = btrfs_next_leaf(extent_root, &path);
2020                         if (ret < 0)
2021                                 goto fail;
2022                         if (ret > 0)
2023                                 break;
2024                         leaf = path.nodes[0];
2025                 }
2026
2027                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2028                 if (key.objectid < cur_byte ||
2029                     key.type != BTRFS_EXTENT_ITEM_KEY) {
2030                         path.slots[0]++;
2031                         goto next;
2032                 }
2033                 if (key.objectid >= end_byte)
2034                         break;
2035
2036                 num_extents++;
2037
2038                 cur_byte = key.objectid;
2039                 num_bytes = key.offset;
2040                 ei = btrfs_item_ptr(leaf, path.slots[0],
2041                                     struct btrfs_extent_item);
2042                 BUG_ON(!(btrfs_extent_flags(leaf, ei) &
2043                          BTRFS_EXTENT_FLAG_DATA));
2044
2045                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2046                 end = ptr + btrfs_item_size_nr(leaf, path.slots[0]);
2047
2048                 ptr += sizeof(struct btrfs_extent_item);
2049
2050                 while (ptr < end) {
2051                         iref = (struct btrfs_extent_inline_ref *)ptr;
2052                         key.type = btrfs_extent_inline_ref_type(leaf, iref);
2053                         BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
2054                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2055                         ref_root = btrfs_extent_data_ref_root(leaf, dref);
2056                         extent_key.objectid =
2057                                 btrfs_extent_data_ref_objectid(leaf, dref);
2058                         extent_key.offset =
2059                                 btrfs_extent_data_ref_offset(leaf, dref);
2060                         extent_key.type = BTRFS_EXTENT_DATA_KEY;
2061                         BUG_ON(btrfs_extent_data_ref_count(leaf, dref) != 1);
2062
2063                         if (ref_root == cur_root->root_key.objectid)
2064                                 break;
2065
2066                         ptr += btrfs_extent_inline_ref_size(key.type);
2067                 }
2068
2069                 if (ptr >= end) {
2070                         path.slots[0]++;
2071                         goto next;
2072                 }
2073
2074                 ret = relocate_one_reference(trans, cur_root, cur_byte,
2075                                              num_bytes, &extent_key,
2076                                              &reloc_tree);
2077                 if (ret < 0)
2078                         goto fail;
2079
2080                 cur_byte += num_bytes;
2081                 btrfs_release_path(&path);
2082
2083                 if (trans->blocks_used >= 4096) {
2084                         ret = btrfs_commit_transaction(trans, cur_root);
2085                         BUG_ON(ret);
2086                         trans = btrfs_start_transaction(cur_root, 1);
2087                         BUG_ON(!trans);
2088                 }
2089         }
2090         btrfs_release_path(&path);
2091
2092         ret = btrfs_commit_transaction(trans, cur_root);
2093         BUG_ON(ret);
2094
2095         if (num_extents > 0 && pass++ < 16)
2096                 goto again;
2097
2098         ret = (num_extents > 0) ? -1 : 0;
2099 fail:
2100         btrfs_release_path(&path);
2101         extent_io_tree_cleanup(&reloc_tree);
2102         return ret;
2103 }
2104
2105 /*
2106  * relocate data in system chunk
2107  */
2108 static int cleanup_sys_chunk(struct btrfs_root *fs_root,
2109                              struct btrfs_root *ext2_root)
2110 {
2111         struct btrfs_block_group_cache *cache;
2112         int i, ret = 0;
2113         u64 offset = 0;
2114         u64 end_byte;
2115
2116         while(1) {
2117                 cache = btrfs_lookup_block_group(fs_root->fs_info, offset);
2118                 if (!cache)
2119                         break;
2120
2121                 end_byte = cache->key.objectid + cache->key.offset;
2122                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2123                         ret = relocate_extents_range(fs_root, ext2_root,
2124                                                      cache->key.objectid,
2125                                                      end_byte);
2126                         if (ret)
2127                                 goto fail;
2128                 }
2129                 offset = end_byte;
2130         }
2131         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2132                 offset = btrfs_sb_offset(i);
2133                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
2134
2135                 ret = relocate_extents_range(fs_root, ext2_root,
2136                                              offset, offset + BTRFS_STRIPE_LEN);
2137                 if (ret)
2138                         goto fail;
2139         }
2140         ret = 0;
2141 fail:
2142         return ret;
2143 }
2144
2145 static int fixup_chunk_mapping(struct btrfs_root *root)
2146 {
2147         struct btrfs_trans_handle *trans;
2148         struct btrfs_fs_info *info = root->fs_info;
2149         struct btrfs_root *chunk_root = info->chunk_root;
2150         struct extent_buffer *leaf;
2151         struct btrfs_key key;
2152         struct btrfs_path path;
2153         struct btrfs_chunk chunk;
2154         unsigned long ptr;
2155         u32 size;
2156         u64 type;
2157         int ret;
2158
2159         btrfs_init_path(&path);
2160
2161         trans = btrfs_start_transaction(root, 1);
2162         BUG_ON(!trans);
2163
2164         /*
2165          * recow the whole chunk tree. this will move all chunk tree blocks
2166          * into system block group.
2167          */
2168         memset(&key, 0, sizeof(key));
2169         while (1) {
2170                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2171                 if (ret < 0)
2172                         goto err;
2173
2174                 ret = btrfs_next_leaf(chunk_root, &path);
2175                 if (ret < 0)
2176                         goto err;
2177                 if (ret > 0)
2178                         break;
2179
2180                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2181                 btrfs_release_path(&path);
2182         }
2183         btrfs_release_path(&path);
2184
2185         /* fixup the system chunk array in super block */
2186         btrfs_set_super_sys_array_size(info->super_copy, 0);
2187
2188         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2189         key.offset = 0;
2190         key.type = BTRFS_CHUNK_ITEM_KEY;
2191
2192         ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 0);
2193         if (ret < 0)
2194                 goto err;
2195         BUG_ON(ret != 0);
2196         while(1) {
2197                 leaf = path.nodes[0];
2198                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2199                         ret = btrfs_next_leaf(chunk_root, &path);
2200                         if (ret < 0)
2201                                 goto err;
2202                         if (ret > 0)
2203                                 break;
2204                         leaf = path.nodes[0];
2205                 }
2206                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2207                 if (key.type != BTRFS_CHUNK_ITEM_KEY)
2208                         goto next;
2209
2210                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2211                 size = btrfs_item_size_nr(leaf, path.slots[0]);
2212                 BUG_ON(size != sizeof(chunk));
2213                 read_extent_buffer(leaf, &chunk, ptr, size);
2214                 type = btrfs_stack_chunk_type(&chunk);
2215
2216                 if (!(type & BTRFS_BLOCK_GROUP_SYSTEM))
2217                         goto next;
2218
2219                 ret = btrfs_add_system_chunk(trans, chunk_root, &key,
2220                                              &chunk, size);
2221                 if (ret)
2222                         goto err;
2223 next:
2224                 path.slots[0]++;
2225         }
2226
2227         ret = btrfs_commit_transaction(trans, root);
2228         BUG_ON(ret);
2229 err:
2230         btrfs_release_path(&path);
2231         return ret;
2232 }
2233
2234 static int do_convert(const char *devname, int datacsum, int packing, int noxattr,
2235                int copylabel, const char *fslabel, int progress)
2236 {
2237         int i, ret;
2238         int fd = -1;
2239         u32 blocksize;
2240         u64 blocks[7];
2241         u64 total_bytes;
2242         u64 super_bytenr;
2243         ext2_filsys ext2_fs;
2244         struct btrfs_root *root;
2245         struct btrfs_root *ext2_root;
2246         struct task_ctx ctx;
2247
2248         ret = open_ext2fs(devname, &ext2_fs);
2249         if (ret) {
2250                 fprintf(stderr, "unable to open the Ext2fs\n");
2251                 goto fail;
2252         }
2253         blocksize = ext2_fs->blocksize;
2254         total_bytes = (u64)ext2_fs->super->s_blocks_count * blocksize;
2255         if (blocksize < 4096) {
2256                 fprintf(stderr, "block size is too small\n");
2257                 goto fail;
2258         }
2259         if (!(ext2_fs->super->s_feature_incompat &
2260               EXT2_FEATURE_INCOMPAT_FILETYPE)) {
2261                 fprintf(stderr, "filetype feature is missing\n");
2262                 goto fail;
2263         }
2264         for (i = 0; i < 7; i++) {
2265                 ret = ext2_alloc_block(ext2_fs, 0, blocks + i);
2266                 if (ret) {
2267                         fprintf(stderr, "not enough free space\n");
2268                         goto fail;
2269                 }
2270                 blocks[i] *= blocksize;
2271         }
2272         super_bytenr = blocks[0];
2273         fd = open(devname, O_RDWR);
2274         if (fd < 0) {
2275                 fprintf(stderr, "unable to open %s\n", devname);
2276                 goto fail;
2277         }
2278         ret = make_btrfs(fd, devname, ext2_fs->super->s_volume_name,
2279                          NULL, blocks, total_bytes, blocksize, blocksize,
2280                          blocksize, blocksize, 0);
2281         if (ret) {
2282                 fprintf(stderr, "unable to create initial ctree: %s\n",
2283                         strerror(-ret));
2284                 goto fail;
2285         }
2286         /* create a system chunk that maps the whole device */
2287         ret = prepare_system_chunk(fd, super_bytenr, blocksize);
2288         if (ret) {
2289                 fprintf(stderr, "unable to update system chunk\n");
2290                 goto fail;
2291         }
2292         root = open_ctree_fd(fd, devname, super_bytenr, OPEN_CTREE_WRITES);
2293         if (!root) {
2294                 fprintf(stderr, "unable to open ctree\n");
2295                 goto fail;
2296         }
2297         ret = cache_free_extents(root, ext2_fs);
2298         if (ret) {
2299                 fprintf(stderr, "error during cache_free_extents %d\n", ret);
2300                 goto fail;
2301         }
2302         root->fs_info->extent_ops = &extent_ops;
2303         /* recover block allocation bitmap */
2304         for (i = 0; i < 7; i++) {
2305                 blocks[i] /= blocksize;
2306                 ext2_free_block(ext2_fs, blocks[i]);
2307         }
2308         ret = init_btrfs(root);
2309         if (ret) {
2310                 fprintf(stderr, "unable to setup the root tree\n");
2311                 goto fail;
2312         }
2313         printf("creating btrfs metadata.\n");
2314         ctx.max_copy_inodes = (ext2_fs->super->s_inodes_count
2315                         - ext2_fs->super->s_free_inodes_count);
2316         ctx.cur_copy_inodes = 0;
2317
2318         if (progress) {
2319                 ctx.info = task_init(print_copied_inodes, after_copied_inodes, &ctx);
2320                 task_start(ctx.info);
2321         }
2322         ret = copy_inodes(root, ext2_fs, datacsum, packing, noxattr, &ctx);
2323         if (ret) {
2324                 fprintf(stderr, "error during copy_inodes %d\n", ret);
2325                 goto fail;
2326         }
2327         if (progress) {
2328                 task_stop(ctx.info);
2329                 task_deinit(ctx.info);
2330         }
2331         printf("creating ext2fs image file.\n");
2332         ext2_root = link_subvol(root, "ext2_saved", EXT2_IMAGE_SUBVOL_OBJECTID);
2333         if (!ext2_root) {
2334                 fprintf(stderr, "unable to create subvol\n");
2335                 goto fail;
2336         }
2337         ret = create_ext2_image(ext2_root, ext2_fs, "image");
2338         if (ret) {
2339                 fprintf(stderr, "error during create_ext2_image %d\n", ret);
2340                 goto fail;
2341         }
2342         memset(root->fs_info->super_copy->label, 0, BTRFS_LABEL_SIZE);
2343         if (copylabel == 1) {
2344                 strncpy(root->fs_info->super_copy->label,
2345                                 ext2_fs->super->s_volume_name, 16);
2346                 fprintf(stderr, "copy label '%s'\n",
2347                                 root->fs_info->super_copy->label);
2348         } else if (copylabel == -1) {
2349                 strncpy(root->fs_info->super_copy->label, fslabel, BTRFS_LABEL_SIZE);
2350                 fprintf(stderr, "set label to '%s'\n", fslabel);
2351         }
2352
2353         printf("cleaning up system chunk.\n");
2354         ret = cleanup_sys_chunk(root, ext2_root);
2355         if (ret) {
2356                 fprintf(stderr, "error during cleanup_sys_chunk %d\n", ret);
2357                 goto fail;
2358         }
2359         ret = close_ctree(root);
2360         if (ret) {
2361                 fprintf(stderr, "error during close_ctree %d\n", ret);
2362                 goto fail;
2363         }
2364         close_ext2fs(ext2_fs);
2365
2366         /*
2367          * If this step succeed, we get a mountable btrfs. Otherwise
2368          * the ext2fs is left unchanged.
2369          */
2370         ret = migrate_super_block(fd, super_bytenr, blocksize);
2371         if (ret) {
2372                 fprintf(stderr, "unable to migrate super block\n");
2373                 goto fail;
2374         }
2375
2376         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2377         if (!root) {
2378                 fprintf(stderr, "unable to open ctree\n");
2379                 goto fail;
2380         }
2381         /* move chunk tree into system chunk. */
2382         ret = fixup_chunk_mapping(root);
2383         if (ret) {
2384                 fprintf(stderr, "error during fixup_chunk_tree\n");
2385                 goto fail;
2386         }
2387         ret = close_ctree(root);
2388         close(fd);
2389
2390         printf("conversion complete.\n");
2391         return 0;
2392 fail:
2393         if (fd != -1)
2394                 close(fd);
2395         fprintf(stderr, "conversion aborted.\n");
2396         return -1;
2397 }
2398
2399 static int may_rollback(struct btrfs_root *root)
2400 {
2401         struct btrfs_fs_info *info = root->fs_info;
2402         struct btrfs_multi_bio *multi = NULL;
2403         u64 bytenr;
2404         u64 length;
2405         u64 physical;
2406         u64 total_bytes;
2407         int num_stripes;
2408         int ret;
2409
2410         if (btrfs_super_num_devices(info->super_copy) != 1)
2411                 goto fail;
2412
2413         bytenr = BTRFS_SUPER_INFO_OFFSET;
2414         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
2415
2416         while (1) {
2417                 ret = btrfs_map_block(&info->mapping_tree, WRITE, bytenr,
2418                                       &length, &multi, 0, NULL);
2419                 if (ret) {
2420                         if (ret == -ENOENT) {
2421                                 /* removed block group at the tail */
2422                                 if (length == (u64)-1)
2423                                         break;
2424
2425                                 /* removed block group in the middle */
2426                                 goto next;
2427                         }
2428                         goto fail;
2429                 }
2430
2431                 num_stripes = multi->num_stripes;
2432                 physical = multi->stripes[0].physical;
2433                 kfree(multi);
2434
2435                 if (num_stripes != 1 || physical != bytenr)
2436                         goto fail;
2437 next:
2438                 bytenr += length;
2439                 if (bytenr >= total_bytes)
2440                         break;
2441         }
2442         return 0;
2443 fail:
2444         return -1;
2445 }
2446
2447 static int do_rollback(const char *devname)
2448 {
2449         int fd = -1;
2450         int ret;
2451         int i;
2452         struct btrfs_root *root;
2453         struct btrfs_root *ext2_root;
2454         struct btrfs_root *chunk_root;
2455         struct btrfs_dir_item *dir;
2456         struct btrfs_inode_item *inode;
2457         struct btrfs_file_extent_item *fi;
2458         struct btrfs_trans_handle *trans;
2459         struct extent_buffer *leaf;
2460         struct btrfs_block_group_cache *cache1;
2461         struct btrfs_block_group_cache *cache2;
2462         struct btrfs_key key;
2463         struct btrfs_path path;
2464         struct extent_io_tree io_tree;
2465         char *buf = NULL;
2466         char *name;
2467         u64 bytenr;
2468         u64 num_bytes;
2469         u64 root_dir;
2470         u64 objectid;
2471         u64 offset;
2472         u64 start;
2473         u64 end;
2474         u64 sb_bytenr;
2475         u64 first_free;
2476         u64 total_bytes;
2477         u32 sectorsize;
2478
2479         extent_io_tree_init(&io_tree);
2480
2481         fd = open(devname, O_RDWR);
2482         if (fd < 0) {
2483                 fprintf(stderr, "unable to open %s\n", devname);
2484                 goto fail;
2485         }
2486         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2487         if (!root) {
2488                 fprintf(stderr, "unable to open ctree\n");
2489                 goto fail;
2490         }
2491         ret = may_rollback(root);
2492         if (ret < 0) {
2493                 fprintf(stderr, "unable to do rollback\n");
2494                 goto fail;
2495         }
2496
2497         sectorsize = root->sectorsize;
2498         buf = malloc(sectorsize);
2499         if (!buf) {
2500                 fprintf(stderr, "unable to allocate memory\n");
2501                 goto fail;
2502         }
2503
2504         btrfs_init_path(&path);
2505
2506         key.objectid = EXT2_IMAGE_SUBVOL_OBJECTID;
2507         key.type = BTRFS_ROOT_ITEM_KEY;
2508         key.offset = (u64)-1;
2509         ext2_root = btrfs_read_fs_root(root->fs_info, &key);
2510         if (!ext2_root || IS_ERR(ext2_root)) {
2511                 fprintf(stderr, "unable to open subvol %llu\n",
2512                         (unsigned long long)key.objectid);
2513                 goto fail;
2514         }
2515
2516         name = "image";
2517         root_dir = btrfs_root_dirid(&root->root_item);
2518         dir = btrfs_lookup_dir_item(NULL, ext2_root, &path,
2519                                    root_dir, name, strlen(name), 0);
2520         if (!dir || IS_ERR(dir)) {
2521                 fprintf(stderr, "unable to find file %s\n", name);
2522                 goto fail;
2523         }
2524         leaf = path.nodes[0];
2525         btrfs_dir_item_key_to_cpu(leaf, dir, &key);
2526         btrfs_release_path(&path);
2527
2528         objectid = key.objectid;
2529
2530         ret = btrfs_lookup_inode(NULL, ext2_root, &path, &key, 0);
2531         if (ret) {
2532                 fprintf(stderr, "unable to find inode item\n");
2533                 goto fail;
2534         }
2535         leaf = path.nodes[0];
2536         inode = btrfs_item_ptr(leaf, path.slots[0], struct btrfs_inode_item);
2537         total_bytes = btrfs_inode_size(leaf, inode);
2538         btrfs_release_path(&path);
2539
2540         key.objectid = objectid;
2541         key.offset = 0;
2542         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
2543         ret = btrfs_search_slot(NULL, ext2_root, &key, &path, 0, 0);
2544         if (ret != 0) {
2545                 fprintf(stderr, "unable to find first file extent\n");
2546                 btrfs_release_path(&path);
2547                 goto fail;
2548         }
2549
2550         /* build mapping tree for the relocated blocks */
2551         for (offset = 0; offset < total_bytes; ) {
2552                 leaf = path.nodes[0];
2553                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2554                         ret = btrfs_next_leaf(root, &path);
2555                         if (ret != 0)
2556                                 break;
2557                         continue;
2558                 }
2559
2560                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2561                 if (key.objectid != objectid || key.offset != offset ||
2562                     btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2563                         break;
2564
2565                 fi = btrfs_item_ptr(leaf, path.slots[0],
2566                                     struct btrfs_file_extent_item);
2567                 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2568                         break;
2569                 if (btrfs_file_extent_compression(leaf, fi) ||
2570                     btrfs_file_extent_encryption(leaf, fi) ||
2571                     btrfs_file_extent_other_encoding(leaf, fi))
2572                         break;
2573
2574                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2575                 /* skip holes and direct mapped extents */
2576                 if (bytenr == 0 || bytenr == offset)
2577                         goto next_extent;
2578
2579                 bytenr += btrfs_file_extent_offset(leaf, fi);
2580                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
2581
2582                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2583                 cache2 =  btrfs_lookup_block_group(root->fs_info,
2584                                                    offset + num_bytes - 1);
2585                 if (!cache1 || cache1 != cache2 ||
2586                     (!(cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
2587                      !intersect_with_sb(offset, num_bytes)))
2588                         break;
2589
2590                 set_extent_bits(&io_tree, offset, offset + num_bytes - 1,
2591                                 EXTENT_LOCKED, GFP_NOFS);
2592                 set_state_private(&io_tree, offset, bytenr);
2593 next_extent:
2594                 offset += btrfs_file_extent_num_bytes(leaf, fi);
2595                 path.slots[0]++;
2596         }
2597         btrfs_release_path(&path);
2598
2599         if (offset < total_bytes) {
2600                 fprintf(stderr, "unable to build extent mapping\n");
2601                 goto fail;
2602         }
2603
2604         first_free = BTRFS_SUPER_INFO_OFFSET + 2 * sectorsize - 1;
2605         first_free &= ~((u64)sectorsize - 1);
2606         /* backup for extent #0 should exist */
2607         if(!test_range_bit(&io_tree, 0, first_free - 1, EXTENT_LOCKED, 1)) {
2608                 fprintf(stderr, "no backup for the first extent\n");
2609                 goto fail;
2610         }
2611         /* force no allocation from system block group */
2612         root->fs_info->system_allocs = -1;
2613         trans = btrfs_start_transaction(root, 1);
2614         BUG_ON(!trans);
2615         /*
2616          * recow the whole chunk tree, this will remove all chunk tree blocks
2617          * from system block group
2618          */
2619         chunk_root = root->fs_info->chunk_root;
2620         memset(&key, 0, sizeof(key));
2621         while (1) {
2622                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2623                 if (ret < 0)
2624                         break;
2625
2626                 ret = btrfs_next_leaf(chunk_root, &path);
2627                 if (ret)
2628                         break;
2629
2630                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2631                 btrfs_release_path(&path);
2632         }
2633         btrfs_release_path(&path);
2634
2635         offset = 0;
2636         num_bytes = 0;
2637         while(1) {
2638                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2639                 if (!cache1)
2640                         break;
2641
2642                 if (cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2643                         num_bytes += btrfs_block_group_used(&cache1->item);
2644
2645                 offset = cache1->key.objectid + cache1->key.offset;
2646         }
2647         /* only extent #0 left in system block group? */
2648         if (num_bytes > first_free) {
2649                 fprintf(stderr, "unable to empty system block group\n");
2650                 goto fail;
2651         }
2652         /* create a system chunk that maps the whole device */
2653         ret = prepare_system_chunk_sb(root->fs_info->super_copy);
2654         if (ret) {
2655                 fprintf(stderr, "unable to update system chunk\n");
2656                 goto fail;
2657         }
2658
2659         ret = btrfs_commit_transaction(trans, root);
2660         BUG_ON(ret);
2661
2662         ret = close_ctree(root);
2663         if (ret) {
2664                 fprintf(stderr, "error during close_ctree %d\n", ret);
2665                 goto fail;
2666         }
2667
2668         /* zero btrfs super block mirrors */
2669         memset(buf, 0, sectorsize);
2670         for (i = 1 ; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2671                 bytenr = btrfs_sb_offset(i);
2672                 if (bytenr >= total_bytes)
2673                         break;
2674                 ret = pwrite(fd, buf, sectorsize, bytenr);
2675                 if (ret != sectorsize) {
2676                         fprintf(stderr,
2677                                 "error during zeroing supreblock %d: %d\n",
2678                                 i, ret);
2679                         goto fail;
2680                 }
2681         }
2682
2683         sb_bytenr = (u64)-1;
2684         /* copy all relocated blocks back */
2685         while(1) {
2686                 ret = find_first_extent_bit(&io_tree, 0, &start, &end,
2687                                             EXTENT_LOCKED);
2688                 if (ret)
2689                         break;
2690
2691                 ret = get_state_private(&io_tree, start, &bytenr);
2692                 BUG_ON(ret);
2693
2694                 clear_extent_bits(&io_tree, start, end, EXTENT_LOCKED,
2695                                   GFP_NOFS);
2696
2697                 while (start <= end) {
2698                         if (start == BTRFS_SUPER_INFO_OFFSET) {
2699                                 sb_bytenr = bytenr;
2700                                 goto next_sector;
2701                         }
2702                         ret = pread(fd, buf, sectorsize, bytenr);
2703                         if (ret < 0) {
2704                                 fprintf(stderr, "error during pread %d\n", ret);
2705                                 goto fail;
2706                         }
2707                         BUG_ON(ret != sectorsize);
2708                         ret = pwrite(fd, buf, sectorsize, start);
2709                         if (ret < 0) {
2710                                 fprintf(stderr, "error during pwrite %d\n", ret);
2711                                 goto fail;
2712                         }
2713                         BUG_ON(ret != sectorsize);
2714 next_sector:
2715                         start += sectorsize;
2716                         bytenr += sectorsize;
2717                 }
2718         }
2719
2720         ret = fsync(fd);
2721         if (ret) {
2722                 fprintf(stderr, "error during fsync %d\n", ret);
2723                 goto fail;
2724         }
2725         /*
2726          * finally, overwrite btrfs super block.
2727          */
2728         ret = pread(fd, buf, sectorsize, sb_bytenr);
2729         if (ret < 0) {
2730                 fprintf(stderr, "error during pread %d\n", ret);
2731                 goto fail;
2732         }
2733         BUG_ON(ret != sectorsize);
2734         ret = pwrite(fd, buf, sectorsize, BTRFS_SUPER_INFO_OFFSET);
2735         if (ret < 0) {
2736                 fprintf(stderr, "error during pwrite %d\n", ret);
2737                 goto fail;
2738         }
2739         BUG_ON(ret != sectorsize);
2740         ret = fsync(fd);
2741         if (ret) {
2742                 fprintf(stderr, "error during fsync %d\n", ret);
2743                 goto fail;
2744         }
2745
2746         close(fd);
2747         free(buf);
2748         extent_io_tree_cleanup(&io_tree);
2749         printf("rollback complete.\n");
2750         return 0;
2751
2752 fail:
2753         if (fd != -1)
2754                 close(fd);
2755         free(buf);
2756         fprintf(stderr, "rollback aborted.\n");
2757         return -1;
2758 }
2759
2760 static void print_usage(void)
2761 {
2762         printf("usage: btrfs-convert [-d] [-i] [-n] [-r] [-l label] [-L] device\n");
2763         printf("\t-d           disable data checksum\n");
2764         printf("\t-i           ignore xattrs and ACLs\n");
2765         printf("\t-n           disable packing of small files\n");
2766         printf("\t-r           roll back to ext2fs\n");
2767         printf("\t-l LABEL     set filesystem label\n");
2768         printf("\t-L           use label from converted fs\n");
2769         printf("\t-p           show converting progress\n");
2770 }
2771
2772 int main(int argc, char *argv[])
2773 {
2774         int ret;
2775         int packing = 1;
2776         int noxattr = 0;
2777         int datacsum = 1;
2778         int rollback = 0;
2779         int copylabel = 0;
2780         int usage_error = 0;
2781         int progress = 0;
2782         char *file;
2783         char *fslabel = NULL;
2784
2785         while(1) {
2786                 int c = getopt(argc, argv, "dinrl:Lp");
2787                 if (c < 0)
2788                         break;
2789                 switch(c) {
2790                         case 'd':
2791                                 datacsum = 0;
2792                                 break;
2793                         case 'i':
2794                                 noxattr = 1;
2795                                 break;
2796                         case 'n':
2797                                 packing = 0;
2798                                 break;
2799                         case 'r':
2800                                 rollback = 1;
2801                                 break;
2802                         case 'l':
2803                                 copylabel = -1;
2804                                 fslabel = strdup(optarg);
2805                                 if (strlen(fslabel) > BTRFS_LABEL_SIZE) {
2806                                         fprintf(stderr,
2807                                                 "warning: label too long, trimmed to %d bytes\n",
2808                                                 BTRFS_LABEL_SIZE);
2809                                         fslabel[BTRFS_LABEL_SIZE] = 0;
2810                                 }
2811                                 break;
2812                         case 'L':
2813                                 copylabel = 1;
2814                                 break;
2815                         case 'p':
2816                                 progress = 1;
2817                                 break;
2818                         default:
2819                                 print_usage();
2820                                 return 1;
2821                 }
2822         }
2823         argc = argc - optind;
2824         set_argv0(argv);
2825         if (check_argc_exact(argc, 1)) {
2826                 print_usage();
2827                 return 1;
2828         }
2829
2830         if (rollback && (!datacsum || noxattr || !packing)) {
2831                 fprintf(stderr,
2832                         "Usage error: -d, -i, -n options do not apply to rollback\n");
2833                 usage_error++;
2834         }
2835
2836         if (usage_error) {
2837                 print_usage();
2838                 return 1;
2839         }
2840
2841         file = argv[optind];
2842         ret = check_mounted(file);
2843         if (ret < 0) {
2844                 fprintf(stderr, "Could not check mount status: %s\n",
2845                         strerror(-ret));
2846                 return 1;
2847         } else if (ret) {
2848                 fprintf(stderr, "%s is mounted\n", file);
2849                 return 1;
2850         }
2851
2852         if (rollback) {
2853                 ret = do_rollback(file);
2854         } else {
2855                 ret = do_convert(file, datacsum, packing, noxattr, copylabel, fslabel, progress);
2856         }
2857         if (ret)
2858                 return 1;
2859         return 0;
2860 }