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 <fcntl.h>
  28 #include <unistd.h>
  29 #include <uuid/uuid.h>
  30 #include <linux/limits.h>
  31 #include <getopt.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 CONV_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_alloc_block_range(ext2_filsys fs, u64 goal, int num,
 144                 u64 *block_ret)
 145 {
 146         blk_t block;
 147         ext2fs_block_bitmap bitmap = fs->block_map;
 148         blk_t start = ext2fs_get_block_bitmap_start(bitmap);
 149         blk_t end = ext2fs_get_block_bitmap_end(bitmap);
 150
 151         for (block = max_t(u64, goal, start); block + num < end; block++) {
 152                 if (ext2fs_fast_test_block_bitmap_range(bitmap, block, num)) {
 153                         ext2fs_fast_mark_block_bitmap_range(bitmap, block,
 154                                         num);
 155                         *block_ret = block;
 156                         return 0;
 157                 }
 158         }
 159         return -ENOSPC;
 160 }
 161
 162 static int ext2_free_block(ext2_filsys fs, u64 block)
 163 {
 164         BUG_ON(block != (blk_t)block);
 165         ext2fs_fast_unmark_block_bitmap(fs->block_map, block);
 166         return 0;
 167 }
 168
 169 static int ext2_free_block_range(ext2_filsys fs, u64 block, int num)
 170 {
 171         BUG_ON(block != (blk_t)block);
 172         ext2fs_fast_unmark_block_bitmap_range(fs->block_map, block, num);
 173         return 0;
 174 }
 175
 176 static int cache_free_extents(struct btrfs_root *root, ext2_filsys ext2_fs)
 177
 178 {
 179         int i, ret = 0;
 180         blk_t block;
 181         u64 bytenr;
 182         u64 blocksize = ext2_fs->blocksize;
 183
 184         block = ext2_fs->super->s_first_data_block;
 185         for (; block < ext2_fs->super->s_blocks_count; block++) {
 186                 if (ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
 187                         continue;
 188                 bytenr = block * blocksize;
 189                 ret = set_extent_dirty(&root->fs_info->free_space_cache,
 190                                        bytenr, bytenr + blocksize - 1, 0);
 191                 BUG_ON(ret);
 192         }
 193
 194         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 195                 bytenr = btrfs_sb_offset(i);
 196                 bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 197                 if (bytenr >= blocksize * ext2_fs->super->s_blocks_count)
 198                         break;
 199                 clear_extent_dirty(&root->fs_info->free_space_cache, bytenr,
 200                                    bytenr + BTRFS_STRIPE_LEN - 1, 0);
 201         }
 202
 203         clear_extent_dirty(&root->fs_info->free_space_cache,
 204                            0, BTRFS_SUPER_INFO_OFFSET - 1, 0);
 205
 206         return 0;
 207 }
 208
 209 static int custom_alloc_extent(struct btrfs_root *root, u64 num_bytes,
 210                                u64 hint_byte, struct btrfs_key *ins,
 211                                int metadata)
 212 {
 213         u64 start;
 214         u64 end;
 215         u64 last = hint_byte;
 216         int ret;
 217         int wrapped = 0;
 218         struct btrfs_block_group_cache *cache;
 219
 220         while(1) {
 221                 ret = find_first_extent_bit(&root->fs_info->free_space_cache,
 222                                             last, &start, &end, EXTENT_DIRTY);
 223                 if (ret) {
 224                         if (wrapped++ == 0) {
 225                                 last = 0;
 226                                 continue;
 227                         } else {
 228                                 goto fail;
 229                         }
 230                 }
 231
 232                 start = max(last, start);
 233                 last = end + 1;
 234                 if (last - start < num_bytes)
 235                         continue;
 236
 237                 last = start + num_bytes;
 238                 if (test_range_bit(&root->fs_info->pinned_extents,
 239                                    start, last - 1, EXTENT_DIRTY, 0))
 240                         continue;
 241
 242                 cache = btrfs_lookup_block_group(root->fs_info, start);
 243                 BUG_ON(!cache);
 244                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM ||
 245                     last > cache->key.objectid + cache->key.offset) {
 246                         last = cache->key.objectid + cache->key.offset;
 247                         continue;
 248                 }
 249
 250                 if (metadata) {
 251                         BUG_ON(num_bytes != root->nodesize);
 252                         if (check_crossing_stripes(start, num_bytes)) {
 253                                 last = round_down(start + num_bytes,
 254                                                   BTRFS_STRIPE_LEN);
 255                                 continue;
 256                         }
 257                 }
 258                 clear_extent_dirty(&root->fs_info->free_space_cache,
 259                                    start, start + num_bytes - 1, 0);
 260
 261                 ins->objectid = start;
 262                 ins->offset = num_bytes;
 263                 ins->type = BTRFS_EXTENT_ITEM_KEY;
 264                 return 0;
 265         }
 266 fail:
 267         fprintf(stderr, "not enough free space\n");
 268         return -ENOSPC;
 269 }
 270
 271 static int intersect_with_sb(u64 bytenr, u64 num_bytes)
 272 {
 273         int i;
 274         u64 offset;
 275
 276         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 277                 offset = btrfs_sb_offset(i);
 278                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
 279
 280                 if (bytenr < offset + BTRFS_STRIPE_LEN &&
 281                     bytenr + num_bytes > offset)
 282                         return 1;
 283         }
 284         return 0;
 285 }
 286
 287 static int custom_free_extent(struct btrfs_root *root, u64 bytenr,
 288                               u64 num_bytes)
 289 {
 290         return intersect_with_sb(bytenr, num_bytes);
 291 }
 292
 293 static struct btrfs_extent_ops extent_ops = {
 294         .alloc_extent = custom_alloc_extent,
 295         .free_extent = custom_free_extent,
 296 };
 297
 298 static int convert_insert_dirent(struct btrfs_trans_handle *trans,
 299                                  struct btrfs_root *root,
 300                                  const char *name, size_t name_len,
 301                                  u64 dir, u64 objectid,
 302                                  u8 file_type, u64 index_cnt,
 303                                  struct btrfs_inode_item *inode)
 304 {
 305         int ret;
 306         u64 inode_size;
 307         struct btrfs_key location = {
 308                 .objectid = objectid,
 309                 .offset = 0,
 310                 .type = BTRFS_INODE_ITEM_KEY,
 311         };
 312
 313         ret = btrfs_insert_dir_item(trans, root, name, name_len,
 314                                     dir, &location, file_type, index_cnt);
 315         if (ret)
 316                 return ret;
 317         ret = btrfs_insert_inode_ref(trans, root, name, name_len,
 318                                      objectid, dir, index_cnt);
 319         if (ret)
 320                 return ret;
 321         inode_size = btrfs_stack_inode_size(inode) + name_len * 2;
 322         btrfs_set_stack_inode_size(inode, inode_size);
 323
 324         return 0;
 325 }
 326
 327 struct dir_iterate_data {
 328         struct btrfs_trans_handle *trans;
 329         struct btrfs_root *root;
 330         struct btrfs_inode_item *inode;
 331         u64 objectid;
 332         u64 index_cnt;
 333         u64 parent;
 334         int errcode;
 335 };
 336
 337 static u8 filetype_conversion_table[EXT2_FT_MAX] = {
 338         [EXT2_FT_UNKNOWN]       = BTRFS_FT_UNKNOWN,
 339         [EXT2_FT_REG_FILE]      = BTRFS_FT_REG_FILE,
 340         [EXT2_FT_DIR]           = BTRFS_FT_DIR,
 341         [EXT2_FT_CHRDEV]        = BTRFS_FT_CHRDEV,
 342         [EXT2_FT_BLKDEV]        = BTRFS_FT_BLKDEV,
 343         [EXT2_FT_FIFO]          = BTRFS_FT_FIFO,
 344         [EXT2_FT_SOCK]          = BTRFS_FT_SOCK,
 345         [EXT2_FT_SYMLINK]       = BTRFS_FT_SYMLINK,
 346 };
 347
 348 static int dir_iterate_proc(ext2_ino_t dir, int entry,
 349                             struct ext2_dir_entry *dirent,
 350                             int offset, int blocksize,
 351                             char *buf,void *priv_data)
 352 {
 353         int ret;
 354         int file_type;
 355         u64 objectid;
 356         char dotdot[] = "..";
 357         struct dir_iterate_data *idata = (struct dir_iterate_data *)priv_data;
 358         int name_len;
 359
 360         name_len = dirent->name_len & 0xFF;
 361
 362         objectid = dirent->inode + INO_OFFSET;
 363         if (!strncmp(dirent->name, dotdot, name_len)) {
 364                 if (name_len == 2) {
 365                         BUG_ON(idata->parent != 0);
 366                         idata->parent = objectid;
 367                 }
 368                 return 0;
 369         }
 370         if (dirent->inode < EXT2_GOOD_OLD_FIRST_INO)
 371                 return 0;
 372
 373         file_type = dirent->name_len >> 8;
 374         BUG_ON(file_type > EXT2_FT_SYMLINK);
 375
 376         ret = convert_insert_dirent(idata->trans, idata->root, dirent->name,
 377                                     name_len, idata->objectid, objectid,
 378                                     filetype_conversion_table[file_type],
 379                                     idata->index_cnt, idata->inode);
 380         if (ret < 0) {
 381                 idata->errcode = ret;
 382                 return BLOCK_ABORT;
 383         }
 384
 385         idata->index_cnt++;
 386         return 0;
 387 }
 388
 389 static int create_dir_entries(struct btrfs_trans_handle *trans,
 390                               struct btrfs_root *root, u64 objectid,
 391                               struct btrfs_inode_item *btrfs_inode,
 392                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 393 {
 394         int ret;
 395         errcode_t err;
 396         struct dir_iterate_data data = {
 397                 .trans          = trans,
 398                 .root           = root,
 399                 .inode          = btrfs_inode,
 400                 .objectid       = objectid,
 401                 .index_cnt      = 2,
 402                 .parent         = 0,
 403                 .errcode        = 0,
 404         };
 405
 406         err = ext2fs_dir_iterate2(ext2_fs, ext2_ino, 0, NULL,
 407                                   dir_iterate_proc, &data);
 408         if (err)
 409                 goto error;
 410         ret = data.errcode;
 411         if (ret == 0 && data.parent == objectid) {
 412                 ret = btrfs_insert_inode_ref(trans, root, "..", 2,
 413                                              objectid, objectid, 0);
 414         }
 415         return ret;
 416 error:
 417         fprintf(stderr, "ext2fs_dir_iterate2: %s\n", error_message(err));
 418         return -1;
 419 }
 420
 421 static int read_disk_extent(struct btrfs_root *root, u64 bytenr,
 422                             u32 num_bytes, char *buffer)
 423 {
 424         int ret;
 425         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
 426
 427         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, bytenr);
 428         if (ret != num_bytes)
 429                 goto fail;
 430         ret = 0;
 431 fail:
 432         if (ret > 0)
 433                 ret = -1;
 434         return ret;
 435 }
 436
 437 static int csum_disk_extent(struct btrfs_trans_handle *trans,
 438                             struct btrfs_root *root,
 439                             u64 disk_bytenr, u64 num_bytes)
 440 {
 441         u32 blocksize = root->sectorsize;
 442         u64 offset;
 443         char *buffer;
 444         int ret = 0;
 445
 446         buffer = malloc(blocksize);
 447         if (!buffer)
 448                 return -ENOMEM;
 449         for (offset = 0; offset < num_bytes; offset += blocksize) {
 450                 ret = read_disk_extent(root, disk_bytenr + offset,
 451                                         blocksize, buffer);
 452                 if (ret)
 453                         break;
 454                 ret = btrfs_csum_file_block(trans,
 455                                             root->fs_info->csum_root,
 456                                             disk_bytenr + num_bytes,
 457                                             disk_bytenr + offset,
 458                                             buffer, blocksize);
 459                 if (ret)
 460                         break;
 461         }
 462         free(buffer);
 463         return ret;
 464 }
 465
 466 struct blk_iterate_data {
 467         struct btrfs_trans_handle *trans;
 468         struct btrfs_root *root;
 469         struct btrfs_inode_item *inode;
 470         u64 objectid;
 471         u64 first_block;
 472         u64 disk_block;
 473         u64 num_blocks;
 474         u64 boundary;
 475         int checksum;
 476         int errcode;
 477 };
 478
 479 static void init_blk_iterate_data(struct blk_iterate_data *data,
 480                                   struct btrfs_trans_handle *trans,
 481                                   struct btrfs_root *root,
 482                                   struct btrfs_inode_item *inode,
 483                                   u64 objectid, int checksum)
 484 {
 485         data->trans             = trans;
 486         data->root              = root;
 487         data->inode             = inode;
 488         data->objectid          = objectid;
 489         data->first_block       = 0;
 490         data->disk_block        = 0;
 491         data->num_blocks        = 0;
 492         data->boundary          = (u64)-1;
 493         data->checksum          = checksum;
 494         data->errcode           = 0;
 495 }
 496
 497 static int record_file_blocks(struct blk_iterate_data *data,
 498                               u64 file_block, u64 disk_block, u64 num_blocks)
 499 {
 500         int ret;
 501         struct btrfs_root *root = data->root;
 502         u64 file_pos = file_block * root->sectorsize;
 503         u64 disk_bytenr = disk_block * root->sectorsize;
 504         u64 num_bytes = num_blocks * root->sectorsize;
 505         ret = btrfs_record_file_extent(data->trans, data->root,
 506                                        data->objectid, data->inode, file_pos,
 507                                        disk_bytenr, num_bytes);
 508
 509         if (ret || !data->checksum || disk_bytenr == 0)
 510                 return ret;
 511
 512         return csum_disk_extent(data->trans, data->root, disk_bytenr,
 513                                 num_bytes);
 514 }
 515
 516 static int block_iterate_proc(u64 disk_block, u64 file_block,
 517                               struct blk_iterate_data *idata)
 518 {
 519         int ret = 0;
 520         int sb_region;
 521         int do_barrier;
 522         struct btrfs_root *root = idata->root;
 523         struct btrfs_block_group_cache *cache;
 524         u64 bytenr = disk_block * root->sectorsize;
 525
 526         sb_region = intersect_with_sb(bytenr, root->sectorsize);
 527         do_barrier = sb_region || disk_block >= idata->boundary;
 528         if ((idata->num_blocks > 0 && do_barrier) ||
 529             (file_block > idata->first_block + idata->num_blocks) ||
 530             (disk_block != idata->disk_block + idata->num_blocks)) {
 531                 if (idata->num_blocks > 0) {
 532                         ret = record_file_blocks(idata, idata->first_block,
 533                                                  idata->disk_block,
 534                                                  idata->num_blocks);
 535                         if (ret)
 536                                 goto fail;
 537                         idata->first_block += idata->num_blocks;
 538                         idata->num_blocks = 0;
 539                 }
 540                 if (file_block > idata->first_block) {
 541                         ret = record_file_blocks(idata, idata->first_block,
 542                                         0, file_block - idata->first_block);
 543                         if (ret)
 544                                 goto fail;
 545                 }
 546
 547                 if (sb_region) {
 548                         bytenr += BTRFS_STRIPE_LEN - 1;
 549                         bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 550                 } else {
 551                         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
 552                         BUG_ON(!cache);
 553                         bytenr = cache->key.objectid + cache->key.offset;
 554                 }
 555
 556                 idata->first_block = file_block;
 557                 idata->disk_block = disk_block;
 558                 idata->boundary = bytenr / root->sectorsize;
 559         }
 560         idata->num_blocks++;
 561 fail:
 562         return ret;
 563 }
 564
 565 static int __block_iterate_proc(ext2_filsys fs, blk_t *blocknr,
 566                                 e2_blkcnt_t blockcnt, blk_t ref_block,
 567                                 int ref_offset, void *priv_data)
 568 {
 569         int ret;
 570         struct blk_iterate_data *idata;
 571         idata = (struct blk_iterate_data *)priv_data;
 572         ret = block_iterate_proc(*blocknr, blockcnt, idata);
 573         if (ret) {
 574                 idata->errcode = ret;
 575                 return BLOCK_ABORT;
 576         }
 577         return 0;
 578 }
 579
 580 /*
 581  * traverse file's data blocks, record these data blocks as file extents.
 582  */
 583 static int create_file_extents(struct btrfs_trans_handle *trans,
 584                                struct btrfs_root *root, u64 objectid,
 585                                struct btrfs_inode_item *btrfs_inode,
 586                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 587                                int datacsum, int packing)
 588 {
 589         int ret;
 590         char *buffer = NULL;
 591         errcode_t err;
 592         u32 last_block;
 593         u32 sectorsize = root->sectorsize;
 594         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 595         struct blk_iterate_data data;
 596
 597         init_blk_iterate_data(&data, trans, root, btrfs_inode, objectid,
 598                               datacsum);
 599
 600         err = ext2fs_block_iterate2(ext2_fs, ext2_ino, BLOCK_FLAG_DATA_ONLY,
 601                                     NULL, __block_iterate_proc, &data);
 602         if (err)
 603                 goto error;
 604         ret = data.errcode;
 605         if (ret)
 606                 goto fail;
 607         if (packing && data.first_block == 0 && data.num_blocks > 0 &&
 608             inode_size <= BTRFS_MAX_INLINE_DATA_SIZE(root)) {
 609                 u64 num_bytes = data.num_blocks * sectorsize;
 610                 u64 disk_bytenr = data.disk_block * sectorsize;
 611                 u64 nbytes;
 612
 613                 buffer = malloc(num_bytes);
 614                 if (!buffer)
 615                         return -ENOMEM;
 616                 ret = read_disk_extent(root, disk_bytenr, num_bytes, buffer);
 617                 if (ret)
 618                         goto fail;
 619                 if (num_bytes > inode_size)
 620                         num_bytes = inode_size;
 621                 ret = btrfs_insert_inline_extent(trans, root, objectid,
 622                                                  0, buffer, num_bytes);
 623                 if (ret)
 624                         goto fail;
 625                 nbytes = btrfs_stack_inode_nbytes(btrfs_inode) + num_bytes;
 626                 btrfs_set_stack_inode_nbytes(btrfs_inode, nbytes);
 627         } else if (data.num_blocks > 0) {
 628                 ret = record_file_blocks(&data, data.first_block,
 629                                          data.disk_block, data.num_blocks);
 630                 if (ret)
 631                         goto fail;
 632         }
 633         data.first_block += data.num_blocks;
 634         last_block = (inode_size + sectorsize - 1) / sectorsize;
 635         if (last_block > data.first_block) {
 636                 ret = record_file_blocks(&data, data.first_block, 0,
 637                                          last_block - data.first_block);
 638         }
 639 fail:
 640         free(buffer);
 641         return ret;
 642 error:
 643         fprintf(stderr, "ext2fs_block_iterate2: %s\n", error_message(err));
 644         return -1;
 645 }
 646
 647 static int create_symbol_link(struct btrfs_trans_handle *trans,
 648                               struct btrfs_root *root, u64 objectid,
 649                               struct btrfs_inode_item *btrfs_inode,
 650                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 651                               struct ext2_inode *ext2_inode)
 652 {
 653         int ret;
 654         char *pathname;
 655         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 656         if (ext2fs_inode_data_blocks(ext2_fs, ext2_inode)) {
 657                 btrfs_set_stack_inode_size(btrfs_inode, inode_size + 1);
 658                 ret = create_file_extents(trans, root, objectid, btrfs_inode,
 659                                           ext2_fs, ext2_ino, 1, 1);
 660                 btrfs_set_stack_inode_size(btrfs_inode, inode_size);
 661                 return ret;
 662         }
 663
 664         pathname = (char *)&(ext2_inode->i_block[0]);
 665         BUG_ON(pathname[inode_size] != 0);
 666         ret = btrfs_insert_inline_extent(trans, root, objectid, 0,
 667                                          pathname, inode_size + 1);
 668         btrfs_set_stack_inode_nbytes(btrfs_inode, inode_size + 1);
 669         return ret;
 670 }
 671
 672 /*
 673  * Following xattr/acl related codes are based on codes in
 674  * fs/ext3/xattr.c and fs/ext3/acl.c
 675  */
 676 #define EXT2_XATTR_BHDR(ptr) ((struct ext2_ext_attr_header *)(ptr))
 677 #define EXT2_XATTR_BFIRST(ptr) \
 678         ((struct ext2_ext_attr_entry *)(EXT2_XATTR_BHDR(ptr) + 1))
 679 #define EXT2_XATTR_IHDR(inode) \
 680         ((struct ext2_ext_attr_header *) ((void *)(inode) + \
 681                 EXT2_GOOD_OLD_INODE_SIZE + (inode)->i_extra_isize))
 682 #define EXT2_XATTR_IFIRST(inode) \
 683         ((struct ext2_ext_attr_entry *) ((void *)EXT2_XATTR_IHDR(inode) + \
 684                 sizeof(EXT2_XATTR_IHDR(inode)->h_magic)))
 685
 686 static int ext2_xattr_check_names(struct ext2_ext_attr_entry *entry,
 687                                   const void *end)
 688 {
 689         struct ext2_ext_attr_entry *next;
 690
 691         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 692                 next = EXT2_EXT_ATTR_NEXT(entry);
 693                 if ((void *)next >= end)
 694                         return -EIO;
 695                 entry = next;
 696         }
 697         return 0;
 698 }
 699
 700 static int ext2_xattr_check_block(const char *buf, size_t size)
 701 {
 702         int error;
 703         struct ext2_ext_attr_header *header = EXT2_XATTR_BHDR(buf);
 704
 705         if (header->h_magic != EXT2_EXT_ATTR_MAGIC ||
 706             header->h_blocks != 1)
 707                 return -EIO;
 708         error = ext2_xattr_check_names(EXT2_XATTR_BFIRST(buf), buf + size);
 709         return error;
 710 }
 711
 712 static int ext2_xattr_check_entry(struct ext2_ext_attr_entry *entry,
 713                                   size_t size)
 714 {
 715         size_t value_size = entry->e_value_size;
 716
 717         if (entry->e_value_block != 0 || value_size > size ||
 718             entry->e_value_offs + value_size > size)
 719                 return -EIO;
 720         return 0;
 721 }
 722
 723 #define EXT2_ACL_VERSION        0x0001
 724
 725 /* 23.2.5 acl_tag_t values */
 726
 727 #define ACL_UNDEFINED_TAG       (0x00)
 728 #define ACL_USER_OBJ            (0x01)
 729 #define ACL_USER                (0x02)
 730 #define ACL_GROUP_OBJ           (0x04)
 731 #define ACL_GROUP               (0x08)
 732 #define ACL_MASK                (0x10)
 733 #define ACL_OTHER               (0x20)
 734
 735 /* 23.2.7 ACL qualifier constants */
 736
 737 #define ACL_UNDEFINED_ID        ((id_t)-1)
 738
 739 typedef struct {
 740         __le16          e_tag;
 741         __le16          e_perm;
 742         __le32          e_id;
 743 } ext2_acl_entry;
 744
 745 typedef struct {
 746         __le16          e_tag;
 747         __le16          e_perm;
 748 } ext2_acl_entry_short;
 749
 750 typedef struct {
 751         __le32          a_version;
 752 } ext2_acl_header;
 753
 754 static inline int ext2_acl_count(size_t size)
 755 {
 756         ssize_t s;
 757         size -= sizeof(ext2_acl_header);
 758         s = size - 4 * sizeof(ext2_acl_entry_short);
 759         if (s < 0) {
 760                 if (size % sizeof(ext2_acl_entry_short))
 761                         return -1;
 762                 return size / sizeof(ext2_acl_entry_short);
 763         } else {
 764                 if (s % sizeof(ext2_acl_entry))
 765                         return -1;
 766                 return s / sizeof(ext2_acl_entry) + 4;
 767         }
 768 }
 769
 770 #define ACL_EA_VERSION          0x0002
 771
 772 typedef struct {
 773         __le16          e_tag;
 774         __le16          e_perm;
 775         __le32          e_id;
 776 } acl_ea_entry;
 777
 778 typedef struct {
 779         __le32          a_version;
 780         acl_ea_entry    a_entries[0];
 781 } acl_ea_header;
 782
 783 static inline size_t acl_ea_size(int count)
 784 {
 785         return sizeof(acl_ea_header) + count * sizeof(acl_ea_entry);
 786 }
 787
 788 static int ext2_acl_to_xattr(void *dst, const void *src,
 789                              size_t dst_size, size_t src_size)
 790 {
 791         int i, count;
 792         const void *end = src + src_size;
 793         acl_ea_header *ext_acl = (acl_ea_header *)dst;
 794         acl_ea_entry *dst_entry = ext_acl->a_entries;
 795         ext2_acl_entry *src_entry;
 796
 797         if (src_size < sizeof(ext2_acl_header))
 798                 goto fail;
 799         if (((ext2_acl_header *)src)->a_version !=
 800             cpu_to_le32(EXT2_ACL_VERSION))
 801                 goto fail;
 802         src += sizeof(ext2_acl_header);
 803         count = ext2_acl_count(src_size);
 804         if (count <= 0)
 805                 goto fail;
 806
 807         BUG_ON(dst_size < acl_ea_size(count));
 808         ext_acl->a_version = cpu_to_le32(ACL_EA_VERSION);
 809         for (i = 0; i < count; i++, dst_entry++) {
 810                 src_entry = (ext2_acl_entry *)src;
 811                 if (src + sizeof(ext2_acl_entry_short) > end)
 812                         goto fail;
 813                 dst_entry->e_tag = src_entry->e_tag;
 814                 dst_entry->e_perm = src_entry->e_perm;
 815                 switch (le16_to_cpu(src_entry->e_tag)) {
 816                 case ACL_USER_OBJ:
 817                 case ACL_GROUP_OBJ:
 818                 case ACL_MASK:
 819                 case ACL_OTHER:
 820                         src += sizeof(ext2_acl_entry_short);
 821                         dst_entry->e_id = cpu_to_le32(ACL_UNDEFINED_ID);
 822                         break;
 823                 case ACL_USER:
 824                 case ACL_GROUP:
 825                         src += sizeof(ext2_acl_entry);
 826                         if (src > end)
 827                                 goto fail;
 828                         dst_entry->e_id = src_entry->e_id;
 829                         break;
 830                 default:
 831                         goto fail;
 832                 }
 833         }
 834         if (src != end)
 835                 goto fail;
 836         return 0;
 837 fail:
 838         return -EINVAL;
 839 }
 840
 841 static char *xattr_prefix_table[] = {
 842         [1] =   "user.",
 843         [2] =   "system.posix_acl_access",
 844         [3] =   "system.posix_acl_default",
 845         [4] =   "trusted.",
 846         [6] =   "security.",
 847 };
 848
 849 static int copy_single_xattr(struct btrfs_trans_handle *trans,
 850                              struct btrfs_root *root, u64 objectid,
 851                              struct ext2_ext_attr_entry *entry,
 852                              const void *data, u32 datalen)
 853 {
 854         int ret = 0;
 855         int name_len;
 856         int name_index;
 857         void *databuf = NULL;
 858         char namebuf[XATTR_NAME_MAX + 1];
 859
 860         name_index = entry->e_name_index;
 861         if (name_index >= ARRAY_SIZE(xattr_prefix_table) ||
 862             xattr_prefix_table[name_index] == NULL)
 863                 return -EOPNOTSUPP;
 864         name_len = strlen(xattr_prefix_table[name_index]) +
 865                    entry->e_name_len;
 866         if (name_len >= sizeof(namebuf))
 867                 return -ERANGE;
 868
 869         if (name_index == 2 || name_index == 3) {
 870                 size_t bufsize = acl_ea_size(ext2_acl_count(datalen));
 871                 databuf = malloc(bufsize);
 872                 if (!databuf)
 873                        return -ENOMEM;
 874                 ret = ext2_acl_to_xattr(databuf, data, bufsize, datalen);
 875                 if (ret)
 876                         goto out;
 877                 data = databuf;
 878                 datalen = bufsize;
 879         }
 880         strncpy(namebuf, xattr_prefix_table[name_index], XATTR_NAME_MAX);
 881         strncat(namebuf, EXT2_EXT_ATTR_NAME(entry), entry->e_name_len);
 882         if (name_len + datalen > BTRFS_LEAF_DATA_SIZE(root) -
 883             sizeof(struct btrfs_item) - sizeof(struct btrfs_dir_item)) {
 884                 fprintf(stderr, "skip large xattr on inode %Lu name %.*s\n",
 885                         objectid - INO_OFFSET, name_len, namebuf);
 886                 goto out;
 887         }
 888         ret = btrfs_insert_xattr_item(trans, root, namebuf, name_len,
 889                                       data, datalen, objectid);
 890 out:
 891         free(databuf);
 892         return ret;
 893 }
 894
 895 static int copy_extended_attrs(struct btrfs_trans_handle *trans,
 896                                struct btrfs_root *root, u64 objectid,
 897                                struct btrfs_inode_item *btrfs_inode,
 898                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 899 {
 900         int ret = 0;
 901         int inline_ea = 0;
 902         errcode_t err;
 903         u32 datalen;
 904         u32 block_size = ext2_fs->blocksize;
 905         u32 inode_size = EXT2_INODE_SIZE(ext2_fs->super);
 906         struct ext2_inode_large *ext2_inode;
 907         struct ext2_ext_attr_entry *entry;
 908         void *data;
 909         char *buffer = NULL;
 910         char inode_buf[EXT2_GOOD_OLD_INODE_SIZE];
 911
 912         if (inode_size <= EXT2_GOOD_OLD_INODE_SIZE) {
 913                 ext2_inode = (struct ext2_inode_large *)inode_buf;
 914         } else {
 915                 ext2_inode = (struct ext2_inode_large *)malloc(inode_size);
 916                 if (!ext2_inode)
 917                        return -ENOMEM;
 918         }
 919         err = ext2fs_read_inode_full(ext2_fs, ext2_ino, (void *)ext2_inode,
 920                                      inode_size);
 921         if (err) {
 922                 fprintf(stderr, "ext2fs_read_inode_full: %s\n",
 923                         error_message(err));
 924                 ret = -1;
 925                 goto out;
 926         }
 927
 928         if (ext2_ino > ext2_fs->super->s_first_ino &&
 929             inode_size > EXT2_GOOD_OLD_INODE_SIZE) {
 930                 if (EXT2_GOOD_OLD_INODE_SIZE +
 931                     ext2_inode->i_extra_isize > inode_size) {
 932                         ret = -EIO;
 933                         goto out;
 934                 }
 935                 if (ext2_inode->i_extra_isize != 0 &&
 936                     EXT2_XATTR_IHDR(ext2_inode)->h_magic ==
 937                     EXT2_EXT_ATTR_MAGIC) {
 938                         inline_ea = 1;
 939                 }
 940         }
 941         if (inline_ea) {
 942                 int total;
 943                 void *end = (void *)ext2_inode + inode_size;
 944                 entry = EXT2_XATTR_IFIRST(ext2_inode);
 945                 total = end - (void *)entry;
 946                 ret = ext2_xattr_check_names(entry, end);
 947                 if (ret)
 948                         goto out;
 949                 while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 950                         ret = ext2_xattr_check_entry(entry, total);
 951                         if (ret)
 952                                 goto out;
 953                         data = (void *)EXT2_XATTR_IFIRST(ext2_inode) +
 954                                 entry->e_value_offs;
 955                         datalen = entry->e_value_size;
 956                         ret = copy_single_xattr(trans, root, objectid,
 957                                                 entry, data, datalen);
 958                         if (ret)
 959                                 goto out;
 960                         entry = EXT2_EXT_ATTR_NEXT(entry);
 961                 }
 962         }
 963
 964         if (ext2_inode->i_file_acl == 0)
 965                 goto out;
 966
 967         buffer = malloc(block_size);
 968         if (!buffer) {
 969                 ret = -ENOMEM;
 970                 goto out;
 971         }
 972         err = ext2fs_read_ext_attr(ext2_fs, ext2_inode->i_file_acl, buffer);
 973         if (err) {
 974                 fprintf(stderr, "ext2fs_read_ext_attr: %s\n",
 975                         error_message(err));
 976                 ret = -1;
 977                 goto out;
 978         }
 979         ret = ext2_xattr_check_block(buffer, block_size);
 980         if (ret)
 981                 goto out;
 982
 983         entry = EXT2_XATTR_BFIRST(buffer);
 984         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 985                 ret = ext2_xattr_check_entry(entry, block_size);
 986                 if (ret)
 987                         goto out;
 988                 data = buffer + entry->e_value_offs;
 989                 datalen = entry->e_value_size;
 990                 ret = copy_single_xattr(trans, root, objectid,
 991                                         entry, data, datalen);
 992                 if (ret)
 993                         goto out;
 994                 entry = EXT2_EXT_ATTR_NEXT(entry);
 995         }
 996 out:
 997         free(buffer);
 998         if ((void *)ext2_inode != inode_buf)
 999                 free(ext2_inode);
1000         return ret;
1001 }
1002 #define MINORBITS       20
1003 #define MKDEV(ma, mi)   (((ma) << MINORBITS) | (mi))
1004
1005 static inline dev_t old_decode_dev(u16 val)
1006 {
1007         return MKDEV((val >> 8) & 255, val & 255);
1008 }
1009
1010 static inline dev_t new_decode_dev(u32 dev)
1011 {
1012         unsigned major = (dev & 0xfff00) >> 8;
1013         unsigned minor = (dev & 0xff) | ((dev >> 12) & 0xfff00);
1014         return MKDEV(major, minor);
1015 }
1016
1017 static int copy_inode_item(struct btrfs_inode_item *dst,
1018                            struct ext2_inode *src, u32 blocksize)
1019 {
1020         btrfs_set_stack_inode_generation(dst, 1);
1021         btrfs_set_stack_inode_size(dst, src->i_size);
1022         btrfs_set_stack_inode_nbytes(dst, 0);
1023         btrfs_set_stack_inode_block_group(dst, 0);
1024         btrfs_set_stack_inode_nlink(dst, src->i_links_count);
1025         btrfs_set_stack_inode_uid(dst, src->i_uid | (src->i_uid_high << 16));
1026         btrfs_set_stack_inode_gid(dst, src->i_gid | (src->i_gid_high << 16));
1027         btrfs_set_stack_inode_mode(dst, src->i_mode);
1028         btrfs_set_stack_inode_rdev(dst, 0);
1029         btrfs_set_stack_inode_flags(dst, 0);
1030         btrfs_set_stack_timespec_sec(&dst->atime, src->i_atime);
1031         btrfs_set_stack_timespec_nsec(&dst->atime, 0);
1032         btrfs_set_stack_timespec_sec(&dst->ctime, src->i_ctime);
1033         btrfs_set_stack_timespec_nsec(&dst->ctime, 0);
1034         btrfs_set_stack_timespec_sec(&dst->mtime, src->i_mtime);
1035         btrfs_set_stack_timespec_nsec(&dst->mtime, 0);
1036         btrfs_set_stack_timespec_sec(&dst->otime, 0);
1037         btrfs_set_stack_timespec_nsec(&dst->otime, 0);
1038
1039         if (S_ISDIR(src->i_mode)) {
1040                 btrfs_set_stack_inode_size(dst, 0);
1041                 btrfs_set_stack_inode_nlink(dst, 1);
1042         }
1043         if (S_ISREG(src->i_mode)) {
1044                 btrfs_set_stack_inode_size(dst, (u64)src->i_size_high << 32 |
1045                                            (u64)src->i_size);
1046         }
1047         if (!S_ISREG(src->i_mode) && !S_ISDIR(src->i_mode) &&
1048             !S_ISLNK(src->i_mode)) {
1049                 if (src->i_block[0]) {
1050                         btrfs_set_stack_inode_rdev(dst,
1051                                 old_decode_dev(src->i_block[0]));
1052                 } else {
1053                         btrfs_set_stack_inode_rdev(dst,
1054                                 new_decode_dev(src->i_block[1]));
1055                 }
1056         }
1057         return 0;
1058 }
1059
1060 /*
1061  * copy a single inode. do all the required works, such as cloning
1062  * inode item, creating file extents and creating directory entries.
1063  */
1064 static int copy_single_inode(struct btrfs_trans_handle *trans,
1065                              struct btrfs_root *root, u64 objectid,
1066                              ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
1067                              struct ext2_inode *ext2_inode,
1068                              int datacsum, int packing, int noxattr)
1069 {
1070         int ret;
1071         struct btrfs_inode_item btrfs_inode;
1072
1073         if (ext2_inode->i_links_count == 0)
1074                 return 0;
1075
1076         copy_inode_item(&btrfs_inode, ext2_inode, ext2_fs->blocksize);
1077         if (!datacsum && S_ISREG(ext2_inode->i_mode)) {
1078                 u32 flags = btrfs_stack_inode_flags(&btrfs_inode) |
1079                             BTRFS_INODE_NODATASUM;
1080                 btrfs_set_stack_inode_flags(&btrfs_inode, flags);
1081         }
1082
1083         switch (ext2_inode->i_mode & S_IFMT) {
1084         case S_IFREG:
1085                 ret = create_file_extents(trans, root, objectid, &btrfs_inode,
1086                                         ext2_fs, ext2_ino, datacsum, packing);
1087                 break;
1088         case S_IFDIR:
1089                 ret = create_dir_entries(trans, root, objectid, &btrfs_inode,
1090                                          ext2_fs, ext2_ino);
1091                 break;
1092         case S_IFLNK:
1093                 ret = create_symbol_link(trans, root, objectid, &btrfs_inode,
1094                                          ext2_fs, ext2_ino, ext2_inode);
1095                 break;
1096         default:
1097                 ret = 0;
1098                 break;
1099         }
1100         if (ret)
1101                 return ret;
1102
1103         if (!noxattr) {
1104                 ret = copy_extended_attrs(trans, root, objectid, &btrfs_inode,
1105                                           ext2_fs, ext2_ino);
1106                 if (ret)
1107                         return ret;
1108         }
1109         return btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1110 }
1111
1112 static int copy_disk_extent(struct btrfs_root *root, u64 dst_bytenr,
1113                             u64 src_bytenr, u32 num_bytes)
1114 {
1115         int ret;
1116         char *buffer;
1117         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
1118
1119         buffer = malloc(num_bytes);
1120         if (!buffer)
1121                 return -ENOMEM;
1122         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, src_bytenr);
1123         if (ret != num_bytes)
1124                 goto fail;
1125         ret = pwrite(fs_devs->latest_bdev, buffer, num_bytes, dst_bytenr);
1126         if (ret != num_bytes)
1127                 goto fail;
1128         ret = 0;
1129 fail:
1130         free(buffer);
1131         if (ret > 0)
1132                 ret = -1;
1133         return ret;
1134 }
1135 /*
1136  * scan ext2's inode bitmap and copy all used inodes.
1137  */
1138 static int copy_inodes(struct btrfs_root *root, ext2_filsys ext2_fs,
1139                        int datacsum, int packing, int noxattr, struct task_ctx *p)
1140 {
1141         int ret;
1142         errcode_t err;
1143         ext2_inode_scan ext2_scan;
1144         struct ext2_inode ext2_inode;
1145         ext2_ino_t ext2_ino;
1146         u64 objectid;
1147         struct btrfs_trans_handle *trans;
1148
1149         trans = btrfs_start_transaction(root, 1);
1150         if (!trans)
1151                 return -ENOMEM;
1152         err = ext2fs_open_inode_scan(ext2_fs, 0, &ext2_scan);
1153         if (err) {
1154                 fprintf(stderr, "ext2fs_open_inode_scan: %s\n", error_message(err));
1155                 return -1;
1156         }
1157         while (!(err = ext2fs_get_next_inode(ext2_scan, &ext2_ino,
1158                                              &ext2_inode))) {
1159                 /* no more inodes */
1160                 if (ext2_ino == 0)
1161                         break;
1162                 /* skip special inode in ext2fs */
1163                 if (ext2_ino < EXT2_GOOD_OLD_FIRST_INO &&
1164                     ext2_ino != EXT2_ROOT_INO)
1165                         continue;
1166                 objectid = ext2_ino + INO_OFFSET;
1167                 ret = copy_single_inode(trans, root,
1168                                         objectid, ext2_fs, ext2_ino,
1169                                         &ext2_inode, datacsum, packing,
1170                                         noxattr);
1171                 p->cur_copy_inodes++;
1172                 if (ret)
1173                         return ret;
1174                 if (trans->blocks_used >= 4096) {
1175                         ret = btrfs_commit_transaction(trans, root);
1176                         BUG_ON(ret);
1177                         trans = btrfs_start_transaction(root, 1);
1178                         BUG_ON(!trans);
1179                 }
1180         }
1181         if (err) {
1182                 fprintf(stderr, "ext2fs_get_next_inode: %s\n", error_message(err));
1183                 return -1;
1184         }
1185         ret = btrfs_commit_transaction(trans, root);
1186         BUG_ON(ret);
1187
1188         return ret;
1189 }
1190
1191 /*
1192  * Construct a range of ext2fs image file.
1193  * scan block allocation bitmap, find all blocks used by the ext2fs
1194  * in this range and create file extents that point to these blocks.
1195  *
1196  * Note: Before calling the function, no file extent points to blocks
1197  *       in this range
1198  */
1199 static int create_image_file_range(struct btrfs_trans_handle *trans,
1200                                    struct btrfs_root *root, u64 objectid,
1201                                    struct btrfs_inode_item *inode,
1202                                    u64 start_byte, u64 end_byte,
1203                                    ext2_filsys ext2_fs, int datacsum)
1204 {
1205         u32 blocksize = ext2_fs->blocksize;
1206         u32 block = start_byte / blocksize;
1207         u32 last_block = (end_byte + blocksize - 1) / blocksize;
1208         int ret = 0;
1209         struct blk_iterate_data data;
1210
1211         init_blk_iterate_data(&data, trans, root, inode, objectid, datacsum);
1212         data.first_block = block;
1213
1214         for (; start_byte < end_byte; block++, start_byte += blocksize) {
1215                 if (!ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
1216                         continue;
1217                 ret = block_iterate_proc(block, block, &data);
1218                 if (ret < 0)
1219                         goto fail;
1220         }
1221         if (data.num_blocks > 0) {
1222                 ret = record_file_blocks(&data, data.first_block,
1223                                          data.disk_block, data.num_blocks);
1224                 if (ret)
1225                         goto fail;
1226                 data.first_block += data.num_blocks;
1227         }
1228         if (last_block > data.first_block) {
1229                 ret = record_file_blocks(&data, data.first_block, 0,
1230                                          last_block - data.first_block);
1231                 if (ret)
1232                         goto fail;
1233         }
1234 fail:
1235         return ret;
1236 }
1237 /*
1238  * Create the ext2fs image file.
1239  */
1240 static int create_ext2_image(struct btrfs_root *root, ext2_filsys ext2_fs,
1241                              const char *name, int datacsum)
1242 {
1243         int ret;
1244         struct btrfs_key key;
1245         struct btrfs_key location;
1246         struct btrfs_path path;
1247         struct btrfs_inode_item btrfs_inode;
1248         struct btrfs_inode_item *inode_item;
1249         struct extent_buffer *leaf;
1250         struct btrfs_fs_info *fs_info = root->fs_info;
1251         struct btrfs_root *extent_root = fs_info->extent_root;
1252         struct btrfs_trans_handle *trans;
1253         struct btrfs_extent_item *ei;
1254         struct btrfs_extent_inline_ref *iref;
1255         struct btrfs_extent_data_ref *dref;
1256         u64 bytenr;
1257         u64 num_bytes;
1258         u64 objectid;
1259         u64 last_byte;
1260         u64 first_free;
1261         u64 total_bytes;
1262         u64 flags = BTRFS_INODE_READONLY;
1263         u32 sectorsize = root->sectorsize;
1264
1265         total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
1266         first_free =  BTRFS_SUPER_INFO_OFFSET + sectorsize * 2 - 1;
1267         first_free &= ~((u64)sectorsize - 1);
1268         if (!datacsum)
1269                 flags |= BTRFS_INODE_NODATASUM;
1270
1271         memset(&btrfs_inode, 0, sizeof(btrfs_inode));
1272         btrfs_set_stack_inode_generation(&btrfs_inode, 1);
1273         btrfs_set_stack_inode_size(&btrfs_inode, total_bytes);
1274         btrfs_set_stack_inode_nlink(&btrfs_inode, 1);
1275         btrfs_set_stack_inode_nbytes(&btrfs_inode, 0);
1276         btrfs_set_stack_inode_mode(&btrfs_inode, S_IFREG | 0400);
1277         btrfs_set_stack_inode_flags(&btrfs_inode,  flags);
1278         btrfs_init_path(&path);
1279         trans = btrfs_start_transaction(root, 1);
1280         BUG_ON(!trans);
1281
1282         objectid = btrfs_root_dirid(&root->root_item);
1283         ret = btrfs_find_free_objectid(trans, root, objectid, &objectid);
1284         if (ret)
1285                 goto fail;
1286
1287         /*
1288          * copy blocks covered by extent #0 to new positions. extent #0 is
1289          * special, we can't rely on relocate_extents_range to relocate it.
1290          */
1291         for (last_byte = 0; last_byte < first_free; last_byte += sectorsize) {
1292                 ret = custom_alloc_extent(root, sectorsize, 0, &key, 0);
1293                 if (ret)
1294                         goto fail;
1295                 ret = copy_disk_extent(root, key.objectid, last_byte,
1296                                        sectorsize);
1297                 if (ret)
1298                         goto fail;
1299                 ret = btrfs_record_file_extent(trans, root, objectid,
1300                                                &btrfs_inode, last_byte,
1301                                                key.objectid, sectorsize);
1302                 if (ret)
1303                         goto fail;
1304                 if (datacsum) {
1305                         ret = csum_disk_extent(trans, root, key.objectid,
1306                                                sectorsize);
1307                         if (ret)
1308                                 goto fail;
1309                 }
1310         }
1311
1312         while(1) {
1313                 key.objectid = last_byte;
1314                 key.offset = 0;
1315                 btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
1316                 ret = btrfs_search_slot(trans, fs_info->extent_root,
1317                                         &key, &path, 0, 0);
1318                 if (ret < 0)
1319                         goto fail;
1320 next:
1321                 leaf = path.nodes[0];
1322                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1323                         ret = btrfs_next_leaf(extent_root, &path);
1324                         if (ret < 0)
1325                                 goto fail;
1326                         if (ret > 0)
1327                                 break;
1328                         leaf = path.nodes[0];
1329                 }
1330                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1331                 if (last_byte > key.objectid ||
1332                     key.type != BTRFS_EXTENT_ITEM_KEY) {
1333                         path.slots[0]++;
1334                         goto next;
1335                 }
1336
1337                 bytenr = key.objectid;
1338                 num_bytes = key.offset;
1339                 ei = btrfs_item_ptr(leaf, path.slots[0],
1340                                     struct btrfs_extent_item);
1341                 if (!(btrfs_extent_flags(leaf, ei) & BTRFS_EXTENT_FLAG_DATA)) {
1342                         path.slots[0]++;
1343                         goto next;
1344                 }
1345
1346                 BUG_ON(btrfs_item_size_nr(leaf, path.slots[0]) != sizeof(*ei) +
1347                        btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY));
1348
1349                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
1350                 key.type = btrfs_extent_inline_ref_type(leaf, iref);
1351                 BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
1352                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1353                 if (btrfs_extent_data_ref_root(leaf, dref) !=
1354                     BTRFS_FS_TREE_OBJECTID) {
1355                         path.slots[0]++;
1356                         goto next;
1357                 }
1358
1359                 if (bytenr > last_byte) {
1360                         ret = create_image_file_range(trans, root, objectid,
1361                                                       &btrfs_inode, last_byte,
1362                                                       bytenr, ext2_fs,
1363                                                       datacsum);
1364                         if (ret)
1365                                 goto fail;
1366                 }
1367                 ret = btrfs_record_file_extent(trans, root, objectid,
1368                                                &btrfs_inode, bytenr, bytenr,
1369                                                num_bytes);
1370                 if (ret)
1371                         goto fail;
1372                 last_byte = bytenr + num_bytes;
1373                 btrfs_release_path(&path);
1374
1375                 if (trans->blocks_used >= 4096) {
1376                         ret = btrfs_commit_transaction(trans, root);
1377                         BUG_ON(ret);
1378                         trans = btrfs_start_transaction(root, 1);
1379                         BUG_ON(!trans);
1380                 }
1381         }
1382         btrfs_release_path(&path);
1383         if (total_bytes > last_byte) {
1384                 ret = create_image_file_range(trans, root, objectid,
1385                                               &btrfs_inode, last_byte,
1386                                               total_bytes, ext2_fs,
1387                                               datacsum);
1388                 if (ret)
1389                         goto fail;
1390         }
1391
1392         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1393         if (ret)
1394                 goto fail;
1395
1396         location.objectid = objectid;
1397         location.offset = 0;
1398         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1399         ret = btrfs_insert_dir_item(trans, root, name, strlen(name),
1400                                     btrfs_root_dirid(&root->root_item),
1401                                     &location, BTRFS_FT_REG_FILE, objectid);
1402         if (ret)
1403                 goto fail;
1404         ret = btrfs_insert_inode_ref(trans, root, name, strlen(name),
1405                                      objectid,
1406                                      btrfs_root_dirid(&root->root_item),
1407                                      objectid);
1408         if (ret)
1409                 goto fail;
1410         location.objectid = btrfs_root_dirid(&root->root_item);
1411         location.offset = 0;
1412         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1413         ret = btrfs_lookup_inode(trans, root, &path, &location, 1);
1414         if (ret)
1415                 goto fail;
1416         leaf = path.nodes[0];
1417         inode_item = btrfs_item_ptr(leaf, path.slots[0],
1418                                     struct btrfs_inode_item);
1419         btrfs_set_inode_size(leaf, inode_item, strlen(name) * 2 +
1420                              btrfs_inode_size(leaf, inode_item));
1421         btrfs_mark_buffer_dirty(leaf);
1422         btrfs_release_path(&path);
1423         ret = btrfs_commit_transaction(trans, root);
1424         BUG_ON(ret);
1425 fail:
1426         btrfs_release_path(&path);
1427         return ret;
1428 }
1429
1430 static struct btrfs_root * link_subvol(struct btrfs_root *root,
1431                 const char *base, u64 root_objectid)
1432 {
1433         struct btrfs_trans_handle *trans;
1434         struct btrfs_fs_info *fs_info = root->fs_info;
1435         struct btrfs_root *tree_root = fs_info->tree_root;
1436         struct btrfs_root *new_root = NULL;
1437         struct btrfs_path *path;
1438         struct btrfs_inode_item *inode_item;
1439         struct extent_buffer *leaf;
1440         struct btrfs_key key;
1441         u64 dirid = btrfs_root_dirid(&root->root_item);
1442         u64 index = 2;
1443         char buf[BTRFS_NAME_LEN + 1]; /* for snprintf null */
1444         int len;
1445         int i;
1446         int ret;
1447
1448         len = strlen(base);
1449         if (len < 1 || len > BTRFS_NAME_LEN)
1450                 return NULL;
1451
1452         path = btrfs_alloc_path();
1453         BUG_ON(!path);
1454
1455         key.objectid = dirid;
1456         key.type = BTRFS_DIR_INDEX_KEY;
1457         key.offset = (u64)-1;
1458
1459         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1460         BUG_ON(ret <= 0);
1461
1462         if (path->slots[0] > 0) {
1463                 path->slots[0]--;
1464                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1465                 if (key.objectid == dirid && key.type == BTRFS_DIR_INDEX_KEY)
1466                         index = key.offset + 1;
1467         }
1468         btrfs_release_path(path);
1469
1470         trans = btrfs_start_transaction(root, 1);
1471         BUG_ON(!trans);
1472
1473         key.objectid = dirid;
1474         key.offset = 0;
1475         key.type =  BTRFS_INODE_ITEM_KEY;
1476
1477         ret = btrfs_lookup_inode(trans, root, path, &key, 1);
1478         BUG_ON(ret);
1479         leaf = path->nodes[0];
1480         inode_item = btrfs_item_ptr(leaf, path->slots[0],
1481                                     struct btrfs_inode_item);
1482
1483         key.objectid = root_objectid;
1484         key.offset = (u64)-1;
1485         key.type = BTRFS_ROOT_ITEM_KEY;
1486
1487         memcpy(buf, base, len);
1488         for (i = 0; i < 1024; i++) {
1489                 ret = btrfs_insert_dir_item(trans, root, buf, len,
1490                                             dirid, &key, BTRFS_FT_DIR, index);
1491                 if (ret != -EEXIST)
1492                         break;
1493                 len = snprintf(buf, ARRAY_SIZE(buf), "%s%d", base, i);
1494                 if (len < 1 || len > BTRFS_NAME_LEN) {
1495                         ret = -EINVAL;
1496                         break;
1497                 }
1498         }
1499         if (ret)
1500                 goto fail;
1501
1502         btrfs_set_inode_size(leaf, inode_item, len * 2 +
1503                              btrfs_inode_size(leaf, inode_item));
1504         btrfs_mark_buffer_dirty(leaf);
1505         btrfs_release_path(path);
1506
1507         /* add the backref first */
1508         ret = btrfs_add_root_ref(trans, tree_root, root_objectid,
1509                                  BTRFS_ROOT_BACKREF_KEY,
1510                                  root->root_key.objectid,
1511                                  dirid, index, buf, len);
1512         BUG_ON(ret);
1513
1514         /* now add the forward ref */
1515         ret = btrfs_add_root_ref(trans, tree_root, root->root_key.objectid,
1516                                  BTRFS_ROOT_REF_KEY, root_objectid,
1517                                  dirid, index, buf, len);
1518
1519         ret = btrfs_commit_transaction(trans, root);
1520         BUG_ON(ret);
1521
1522         new_root = btrfs_read_fs_root(fs_info, &key);
1523         if (IS_ERR(new_root))
1524                 new_root = NULL;
1525 fail:
1526         btrfs_free_path(path);
1527         return new_root;
1528 }
1529
1530 static int create_chunk_mapping(struct btrfs_trans_handle *trans,
1531                                 struct btrfs_root *root)
1532 {
1533         struct btrfs_fs_info *info = root->fs_info;
1534         struct btrfs_root *chunk_root = info->chunk_root;
1535         struct btrfs_root *extent_root = info->extent_root;
1536         struct btrfs_device *device;
1537         struct btrfs_block_group_cache *cache;
1538         struct btrfs_dev_extent *extent;
1539         struct extent_buffer *leaf;
1540         struct btrfs_chunk chunk;
1541         struct btrfs_key key;
1542         struct btrfs_path path;
1543         u64 cur_start;
1544         u64 total_bytes;
1545         u64 chunk_objectid;
1546         int ret;
1547
1548         btrfs_init_path(&path);
1549
1550         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
1551         chunk_objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
1552
1553         BUG_ON(list_empty(&info->fs_devices->devices));
1554         device = list_entry(info->fs_devices->devices.next,
1555                             struct btrfs_device, dev_list);
1556         BUG_ON(device->devid != info->fs_devices->latest_devid);
1557
1558         /* delete device extent created by make_btrfs */
1559         key.objectid = device->devid;
1560         key.offset = 0;
1561         key.type = BTRFS_DEV_EXTENT_KEY;
1562         ret = btrfs_search_slot(trans, device->dev_root, &key, &path, -1, 1);
1563         if (ret < 0)
1564                 goto err;
1565
1566         BUG_ON(ret > 0);
1567         ret = btrfs_del_item(trans, device->dev_root, &path);
1568         if (ret)
1569                 goto err;
1570         btrfs_release_path(&path);
1571
1572         /* delete chunk item created by make_btrfs */
1573         key.objectid = chunk_objectid;
1574         key.offset = 0;
1575         key.type = BTRFS_CHUNK_ITEM_KEY;
1576         ret = btrfs_search_slot(trans, chunk_root, &key, &path, -1, 1);
1577         if (ret < 0)
1578                 goto err;
1579
1580         BUG_ON(ret > 0);
1581         ret = btrfs_del_item(trans, chunk_root, &path);
1582         if (ret)
1583                 goto err;
1584         btrfs_release_path(&path);
1585
1586         /* for each block group, create device extent and chunk item */
1587         cur_start = 0;
1588         while (cur_start < total_bytes) {
1589                 cache = btrfs_lookup_block_group(root->fs_info, cur_start);
1590                 BUG_ON(!cache);
1591
1592                 /* insert device extent */
1593                 key.objectid = device->devid;
1594                 key.offset = cache->key.objectid;
1595                 key.type = BTRFS_DEV_EXTENT_KEY;
1596                 ret = btrfs_insert_empty_item(trans, device->dev_root, &path,
1597                                               &key, sizeof(*extent));
1598                 if (ret)
1599                         goto err;
1600
1601                 leaf = path.nodes[0];
1602                 extent = btrfs_item_ptr(leaf, path.slots[0],
1603                                         struct btrfs_dev_extent);
1604
1605                 btrfs_set_dev_extent_chunk_tree(leaf, extent,
1606                                                 chunk_root->root_key.objectid);
1607                 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
1608                                                     chunk_objectid);
1609                 btrfs_set_dev_extent_chunk_offset(leaf, extent,
1610                                                   cache->key.objectid);
1611                 btrfs_set_dev_extent_length(leaf, extent, cache->key.offset);
1612                 write_extent_buffer(leaf, root->fs_info->chunk_tree_uuid,
1613                     (unsigned long)btrfs_dev_extent_chunk_tree_uuid(extent),
1614                     BTRFS_UUID_SIZE);
1615                 btrfs_mark_buffer_dirty(leaf);
1616                 btrfs_release_path(&path);
1617
1618                 /* insert chunk item */
1619                 btrfs_set_stack_chunk_length(&chunk, cache->key.offset);
1620                 btrfs_set_stack_chunk_owner(&chunk,
1621                                             extent_root->root_key.objectid);
1622                 btrfs_set_stack_chunk_stripe_len(&chunk, BTRFS_STRIPE_LEN);
1623                 btrfs_set_stack_chunk_type(&chunk, cache->flags);
1624                 btrfs_set_stack_chunk_io_align(&chunk, device->io_align);
1625                 btrfs_set_stack_chunk_io_width(&chunk, device->io_width);
1626                 btrfs_set_stack_chunk_sector_size(&chunk, device->sector_size);
1627                 btrfs_set_stack_chunk_num_stripes(&chunk, 1);
1628                 btrfs_set_stack_chunk_sub_stripes(&chunk, 0);
1629                 btrfs_set_stack_stripe_devid(&chunk.stripe, device->devid);
1630                 btrfs_set_stack_stripe_offset(&chunk.stripe,
1631                                               cache->key.objectid);
1632                 memcpy(&chunk.stripe.dev_uuid, device->uuid, BTRFS_UUID_SIZE);
1633
1634                 key.objectid = chunk_objectid;
1635                 key.offset = cache->key.objectid;
1636                 key.type = BTRFS_CHUNK_ITEM_KEY;
1637
1638                 ret = btrfs_insert_item(trans, chunk_root, &key, &chunk,
1639                                         btrfs_chunk_item_size(1));
1640                 if (ret)
1641                         goto err;
1642
1643                 cur_start = cache->key.objectid + cache->key.offset;
1644         }
1645
1646         device->bytes_used = total_bytes;
1647         ret = btrfs_update_device(trans, device);
1648 err:
1649         btrfs_release_path(&path);
1650         return ret;
1651 }
1652
1653 static int create_subvol(struct btrfs_trans_handle *trans,
1654                          struct btrfs_root *root, u64 root_objectid)
1655 {
1656         struct extent_buffer *tmp;
1657         struct btrfs_root *new_root;
1658         struct btrfs_key key;
1659         struct btrfs_root_item root_item;
1660         int ret;
1661
1662         ret = btrfs_copy_root(trans, root, root->node, &tmp,
1663                               root_objectid);
1664         BUG_ON(ret);
1665
1666         memcpy(&root_item, &root->root_item, sizeof(root_item));
1667         btrfs_set_root_bytenr(&root_item, tmp->start);
1668         btrfs_set_root_level(&root_item, btrfs_header_level(tmp));
1669         btrfs_set_root_generation(&root_item, trans->transid);
1670         free_extent_buffer(tmp);
1671
1672         key.objectid = root_objectid;
1673         key.type = BTRFS_ROOT_ITEM_KEY;
1674         key.offset = trans->transid;
1675         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1676                                 &key, &root_item);
1677
1678         key.offset = (u64)-1;
1679         new_root = btrfs_read_fs_root(root->fs_info, &key);
1680         BUG_ON(!new_root || IS_ERR(new_root));
1681
1682         ret = btrfs_make_root_dir(trans, new_root, BTRFS_FIRST_FREE_OBJECTID);
1683         BUG_ON(ret);
1684
1685         return 0;
1686 }
1687
1688 static int init_btrfs(struct btrfs_root *root)
1689 {
1690         int ret;
1691         struct btrfs_key location;
1692         struct btrfs_trans_handle *trans;
1693         struct btrfs_fs_info *fs_info = root->fs_info;
1694         struct extent_buffer *tmp;
1695
1696         trans = btrfs_start_transaction(root, 1);
1697         BUG_ON(!trans);
1698         ret = btrfs_make_block_groups(trans, root);
1699         if (ret)
1700                 goto err;
1701         ret = btrfs_fix_block_accounting(trans, root);
1702         if (ret)
1703                 goto err;
1704         ret = create_chunk_mapping(trans, root);
1705         if (ret)
1706                 goto err;
1707         ret = btrfs_make_root_dir(trans, fs_info->tree_root,
1708                                   BTRFS_ROOT_TREE_DIR_OBJECTID);
1709         if (ret)
1710                 goto err;
1711         memcpy(&location, &root->root_key, sizeof(location));
1712         location.offset = (u64)-1;
1713         ret = btrfs_insert_dir_item(trans, fs_info->tree_root, "default", 7,
1714                                 btrfs_super_root_dir(fs_info->super_copy),
1715                                 &location, BTRFS_FT_DIR, 0);
1716         if (ret)
1717                 goto err;
1718         ret = btrfs_insert_inode_ref(trans, fs_info->tree_root, "default", 7,
1719                                 location.objectid,
1720                                 btrfs_super_root_dir(fs_info->super_copy), 0);
1721         if (ret)
1722                 goto err;
1723         btrfs_set_root_dirid(&fs_info->fs_root->root_item,
1724                              BTRFS_FIRST_FREE_OBJECTID);
1725
1726         /* subvol for ext2 image file */
1727         ret = create_subvol(trans, root, CONV_IMAGE_SUBVOL_OBJECTID);
1728         BUG_ON(ret);
1729         /* subvol for data relocation */
1730         ret = create_subvol(trans, root, BTRFS_DATA_RELOC_TREE_OBJECTID);
1731         BUG_ON(ret);
1732
1733         extent_buffer_get(fs_info->csum_root->node);
1734         ret = __btrfs_cow_block(trans, fs_info->csum_root,
1735                                 fs_info->csum_root->node, NULL, 0, &tmp, 0, 0);
1736         BUG_ON(ret);
1737         free_extent_buffer(tmp);
1738
1739         ret = btrfs_commit_transaction(trans, root);
1740         BUG_ON(ret);
1741 err:
1742         return ret;
1743 }
1744
1745 /*
1746  * Migrate super block to its default position and zero 0 ~ 16k
1747  */
1748 static int migrate_super_block(int fd, u64 old_bytenr, u32 sectorsize)
1749 {
1750         int ret;
1751         struct extent_buffer *buf;
1752         struct btrfs_super_block *super;
1753         u32 len;
1754         u32 bytenr;
1755
1756         BUG_ON(sectorsize < sizeof(*super));
1757         buf = malloc(sizeof(*buf) + sectorsize);
1758         if (!buf)
1759                 return -ENOMEM;
1760
1761         buf->len = sectorsize;
1762         ret = pread(fd, buf->data, sectorsize, old_bytenr);
1763         if (ret != sectorsize)
1764                 goto fail;
1765
1766         super = (struct btrfs_super_block *)buf->data;
1767         BUG_ON(btrfs_super_bytenr(super) != old_bytenr);
1768         btrfs_set_super_bytenr(super, BTRFS_SUPER_INFO_OFFSET);
1769
1770         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1771         ret = pwrite(fd, buf->data, sectorsize, BTRFS_SUPER_INFO_OFFSET);
1772         if (ret != sectorsize)
1773                 goto fail;
1774
1775         ret = fsync(fd);
1776         if (ret)
1777                 goto fail;
1778
1779         memset(buf->data, 0, sectorsize);
1780         for (bytenr = 0; bytenr < BTRFS_SUPER_INFO_OFFSET; ) {
1781                 len = BTRFS_SUPER_INFO_OFFSET - bytenr;
1782                 if (len > sectorsize)
1783                         len = sectorsize;
1784                 ret = pwrite(fd, buf->data, len, bytenr);
1785                 if (ret != len) {
1786                         fprintf(stderr, "unable to zero fill device\n");
1787                         break;
1788                 }
1789                 bytenr += len;
1790         }
1791         ret = 0;
1792         fsync(fd);
1793 fail:
1794         free(buf);
1795         if (ret > 0)
1796                 ret = -1;
1797         return ret;
1798 }
1799
1800 static int prepare_system_chunk_sb(struct btrfs_super_block *super)
1801 {
1802         struct btrfs_chunk *chunk;
1803         struct btrfs_disk_key *key;
1804         u32 sectorsize = btrfs_super_sectorsize(super);
1805
1806         key = (struct btrfs_disk_key *)(super->sys_chunk_array);
1807         chunk = (struct btrfs_chunk *)(super->sys_chunk_array +
1808                                        sizeof(struct btrfs_disk_key));
1809
1810         btrfs_set_disk_key_objectid(key, BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1811         btrfs_set_disk_key_type(key, BTRFS_CHUNK_ITEM_KEY);
1812         btrfs_set_disk_key_offset(key, 0);
1813
1814         btrfs_set_stack_chunk_length(chunk, btrfs_super_total_bytes(super));
1815         btrfs_set_stack_chunk_owner(chunk, BTRFS_EXTENT_TREE_OBJECTID);
1816         btrfs_set_stack_chunk_stripe_len(chunk, BTRFS_STRIPE_LEN);
1817         btrfs_set_stack_chunk_type(chunk, BTRFS_BLOCK_GROUP_SYSTEM);
1818         btrfs_set_stack_chunk_io_align(chunk, sectorsize);
1819         btrfs_set_stack_chunk_io_width(chunk, sectorsize);
1820         btrfs_set_stack_chunk_sector_size(chunk, sectorsize);
1821         btrfs_set_stack_chunk_num_stripes(chunk, 1);
1822         btrfs_set_stack_chunk_sub_stripes(chunk, 0);
1823         chunk->stripe.devid = super->dev_item.devid;
1824         btrfs_set_stack_stripe_offset(&chunk->stripe, 0);
1825         memcpy(chunk->stripe.dev_uuid, super->dev_item.uuid, BTRFS_UUID_SIZE);
1826         btrfs_set_super_sys_array_size(super, sizeof(*key) + sizeof(*chunk));
1827         return 0;
1828 }
1829
1830 static int prepare_system_chunk(int fd, u64 sb_bytenr)
1831 {
1832         int ret;
1833         struct extent_buffer *buf;
1834         struct btrfs_super_block *super;
1835
1836         BUG_ON(BTRFS_SUPER_INFO_SIZE < sizeof(*super));
1837         buf = malloc(sizeof(*buf) + BTRFS_SUPER_INFO_SIZE);
1838         if (!buf)
1839                 return -ENOMEM;
1840
1841         buf->len = BTRFS_SUPER_INFO_SIZE;
1842         ret = pread(fd, buf->data, BTRFS_SUPER_INFO_SIZE, sb_bytenr);
1843         if (ret != BTRFS_SUPER_INFO_SIZE)
1844                 goto fail;
1845
1846         super = (struct btrfs_super_block *)buf->data;
1847         BUG_ON(btrfs_super_bytenr(super) != sb_bytenr);
1848         BUG_ON(btrfs_super_num_devices(super) != 1);
1849
1850         ret = prepare_system_chunk_sb(super);
1851         if (ret)
1852                 goto fail;
1853
1854         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1855         ret = pwrite(fd, buf->data, BTRFS_SUPER_INFO_SIZE, sb_bytenr);
1856         if (ret != BTRFS_SUPER_INFO_SIZE)
1857                 goto fail;
1858
1859         ret = 0;
1860 fail:
1861         free(buf);
1862         if (ret > 0)
1863                 ret = -1;
1864         return ret;
1865 }
1866
1867 static int relocate_one_reference(struct btrfs_trans_handle *trans,
1868                                   struct btrfs_root *root,
1869                                   u64 extent_start, u64 extent_size,
1870                                   struct btrfs_key *extent_key,
1871                                   struct extent_io_tree *reloc_tree)
1872 {
1873         struct extent_buffer *leaf;
1874         struct btrfs_file_extent_item *fi;
1875         struct btrfs_key key;
1876         struct btrfs_path path;
1877         struct btrfs_inode_item inode;
1878         struct blk_iterate_data data;
1879         u64 bytenr;
1880         u64 num_bytes;
1881         u64 cur_offset;
1882         u64 new_pos;
1883         u64 nbytes;
1884         u64 sector_end;
1885         u32 sectorsize = root->sectorsize;
1886         unsigned long ptr;
1887         int datacsum;
1888         int fd;
1889         int ret;
1890
1891         btrfs_init_path(&path);
1892         ret = btrfs_search_slot(trans, root, extent_key, &path, -1, 1);
1893         if (ret)
1894                 goto fail;
1895
1896         leaf = path.nodes[0];
1897         fi = btrfs_item_ptr(leaf, path.slots[0],
1898                             struct btrfs_file_extent_item);
1899         BUG_ON(btrfs_file_extent_offset(leaf, fi) > 0);
1900         if (extent_start != btrfs_file_extent_disk_bytenr(leaf, fi) ||
1901             extent_size != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1902                 ret = 1;
1903                 goto fail;
1904         }
1905
1906         bytenr = extent_start + btrfs_file_extent_offset(leaf, fi);
1907         num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
1908
1909         ret = btrfs_del_item(trans, root, &path);
1910         if (ret)
1911                 goto fail;
1912
1913         ret = btrfs_free_extent(trans, root, extent_start, extent_size, 0,
1914                                 root->root_key.objectid,
1915                                 extent_key->objectid, extent_key->offset);
1916         if (ret)
1917                 goto fail;
1918
1919         btrfs_release_path(&path);
1920
1921         key.objectid = extent_key->objectid;
1922         key.offset = 0;
1923         key.type =  BTRFS_INODE_ITEM_KEY;
1924         ret = btrfs_lookup_inode(trans, root, &path, &key, 0);
1925         if (ret)
1926                 goto fail;
1927
1928         leaf = path.nodes[0];
1929         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1930         read_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1931         btrfs_release_path(&path);
1932
1933         BUG_ON(num_bytes & (sectorsize - 1));
1934         nbytes = btrfs_stack_inode_nbytes(&inode) - num_bytes;
1935         btrfs_set_stack_inode_nbytes(&inode, nbytes);
1936         datacsum = !(btrfs_stack_inode_flags(&inode) & BTRFS_INODE_NODATASUM);
1937
1938         init_blk_iterate_data(&data, trans, root, &inode, extent_key->objectid,
1939                               datacsum);
1940         data.first_block = extent_key->offset;
1941
1942         cur_offset = extent_key->offset;
1943         while (num_bytes > 0) {
1944                 sector_end = bytenr + sectorsize - 1;
1945                 if (test_range_bit(reloc_tree, bytenr, sector_end,
1946                                    EXTENT_LOCKED, 1)) {
1947                         ret = get_state_private(reloc_tree, bytenr, &new_pos);
1948                         BUG_ON(ret);
1949                 } else {
1950                         ret = custom_alloc_extent(root, sectorsize, 0, &key, 0);
1951                         if (ret)
1952                                 goto fail;
1953                         new_pos = key.objectid;
1954
1955                         if (cur_offset == extent_key->offset) {
1956                                 fd = root->fs_info->fs_devices->latest_bdev;
1957                                 readahead(fd, bytenr, num_bytes);
1958                         }
1959                         ret = copy_disk_extent(root, new_pos, bytenr,
1960                                                sectorsize);
1961                         if (ret)
1962                                 goto fail;
1963                         ret = set_extent_bits(reloc_tree, bytenr, sector_end,
1964                                               EXTENT_LOCKED, GFP_NOFS);
1965                         BUG_ON(ret);
1966                         ret = set_state_private(reloc_tree, bytenr, new_pos);
1967                         BUG_ON(ret);
1968                 }
1969
1970                 ret = block_iterate_proc(new_pos / sectorsize,
1971                                          cur_offset / sectorsize, &data);
1972                 if (ret < 0)
1973                         goto fail;
1974
1975                 cur_offset += sectorsize;
1976                 bytenr += sectorsize;
1977                 num_bytes -= sectorsize;
1978         }
1979
1980         if (data.num_blocks > 0) {
1981                 ret = record_file_blocks(&data, data.first_block,
1982                                          data.disk_block, data.num_blocks);
1983                 if (ret)
1984                         goto fail;
1985         }
1986
1987         key.objectid = extent_key->objectid;
1988         key.offset = 0;
1989         key.type =  BTRFS_INODE_ITEM_KEY;
1990         ret = btrfs_lookup_inode(trans, root, &path, &key, 1);
1991         if (ret)
1992                 goto fail;
1993
1994         leaf = path.nodes[0];
1995         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1996         write_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1997         btrfs_mark_buffer_dirty(leaf);
1998         btrfs_release_path(&path);
1999
2000 fail:
2001         btrfs_release_path(&path);
2002         return ret;
2003 }
2004
2005 static int relocate_extents_range(struct btrfs_root *fs_root,
2006                                   struct btrfs_root *image_root,
2007                                   u64 start_byte, u64 end_byte)
2008 {
2009         struct btrfs_fs_info *info = fs_root->fs_info;
2010         struct btrfs_root *extent_root = info->extent_root;
2011         struct btrfs_root *cur_root = NULL;
2012         struct btrfs_trans_handle *trans;
2013         struct btrfs_extent_data_ref *dref;
2014         struct btrfs_extent_inline_ref *iref;
2015         struct btrfs_extent_item *ei;
2016         struct extent_buffer *leaf;
2017         struct btrfs_key key;
2018         struct btrfs_key extent_key;
2019         struct btrfs_path path;
2020         struct extent_io_tree reloc_tree;
2021         unsigned long ptr;
2022         unsigned long end;
2023         u64 cur_byte;
2024         u64 num_bytes;
2025         u64 ref_root;
2026         u64 num_extents;
2027         int pass = 0;
2028         int ret;
2029
2030         btrfs_init_path(&path);
2031         extent_io_tree_init(&reloc_tree);
2032
2033         key.objectid = start_byte;
2034         key.offset = 0;
2035         key.type = BTRFS_EXTENT_ITEM_KEY;
2036         ret = btrfs_search_slot(NULL, extent_root, &key, &path, 0, 0);
2037         if (ret < 0)
2038                 goto fail;
2039         if (ret > 0) {
2040                 ret = btrfs_previous_item(extent_root, &path, 0,
2041                                           BTRFS_EXTENT_ITEM_KEY);
2042                 if (ret < 0)
2043                         goto fail;
2044                 if (ret == 0) {
2045                         leaf = path.nodes[0];
2046                         btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2047                         if (key.objectid + key.offset > start_byte)
2048                                 start_byte = key.objectid;
2049                 }
2050         }
2051         btrfs_release_path(&path);
2052 again:
2053         cur_root = (pass % 2 == 0) ? image_root : fs_root;
2054         num_extents = 0;
2055
2056         trans = btrfs_start_transaction(cur_root, 1);
2057         BUG_ON(!trans);
2058
2059         cur_byte = start_byte;
2060         while (1) {
2061                 key.objectid = cur_byte;
2062                 key.offset = 0;
2063                 key.type = BTRFS_EXTENT_ITEM_KEY;
2064                 ret = btrfs_search_slot(trans, extent_root,
2065                                         &key, &path, 0, 0);
2066                 if (ret < 0)
2067                         goto fail;
2068 next:
2069                 leaf = path.nodes[0];
2070                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2071                         ret = btrfs_next_leaf(extent_root, &path);
2072                         if (ret < 0)
2073                                 goto fail;
2074                         if (ret > 0)
2075                                 break;
2076                         leaf = path.nodes[0];
2077                 }
2078
2079                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2080                 if (key.objectid < cur_byte ||
2081                     key.type != BTRFS_EXTENT_ITEM_KEY) {
2082                         path.slots[0]++;
2083                         goto next;
2084                 }
2085                 if (key.objectid >= end_byte)
2086                         break;
2087
2088                 num_extents++;
2089
2090                 cur_byte = key.objectid;
2091                 num_bytes = key.offset;
2092                 ei = btrfs_item_ptr(leaf, path.slots[0],
2093                                     struct btrfs_extent_item);
2094                 BUG_ON(!(btrfs_extent_flags(leaf, ei) &
2095                          BTRFS_EXTENT_FLAG_DATA));
2096
2097                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2098                 end = ptr + btrfs_item_size_nr(leaf, path.slots[0]);
2099
2100                 ptr += sizeof(struct btrfs_extent_item);
2101
2102                 while (ptr < end) {
2103                         iref = (struct btrfs_extent_inline_ref *)ptr;
2104                         key.type = btrfs_extent_inline_ref_type(leaf, iref);
2105                         BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
2106                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2107                         ref_root = btrfs_extent_data_ref_root(leaf, dref);
2108                         extent_key.objectid =
2109                                 btrfs_extent_data_ref_objectid(leaf, dref);
2110                         extent_key.offset =
2111                                 btrfs_extent_data_ref_offset(leaf, dref);
2112                         extent_key.type = BTRFS_EXTENT_DATA_KEY;
2113                         BUG_ON(btrfs_extent_data_ref_count(leaf, dref) != 1);
2114
2115                         if (ref_root == cur_root->root_key.objectid)
2116                                 break;
2117
2118                         ptr += btrfs_extent_inline_ref_size(key.type);
2119                 }
2120
2121                 if (ptr >= end) {
2122                         path.slots[0]++;
2123                         goto next;
2124                 }
2125
2126                 ret = relocate_one_reference(trans, cur_root, cur_byte,
2127                                              num_bytes, &extent_key,
2128                                              &reloc_tree);
2129                 if (ret < 0)
2130                         goto fail;
2131
2132                 cur_byte += num_bytes;
2133                 btrfs_release_path(&path);
2134
2135                 if (trans->blocks_used >= 4096) {
2136                         ret = btrfs_commit_transaction(trans, cur_root);
2137                         BUG_ON(ret);
2138                         trans = btrfs_start_transaction(cur_root, 1);
2139                         BUG_ON(!trans);
2140                 }
2141         }
2142         btrfs_release_path(&path);
2143
2144         ret = btrfs_commit_transaction(trans, cur_root);
2145         BUG_ON(ret);
2146
2147         if (num_extents > 0 && pass++ < 16)
2148                 goto again;
2149
2150         ret = (num_extents > 0) ? -1 : 0;
2151 fail:
2152         btrfs_release_path(&path);
2153         extent_io_tree_cleanup(&reloc_tree);
2154         return ret;
2155 }
2156
2157 /*
2158  * relocate data in system chunk
2159  */
2160 static int cleanup_sys_chunk(struct btrfs_root *fs_root,
2161                              struct btrfs_root *image_root)
2162 {
2163         struct btrfs_block_group_cache *cache;
2164         int i, ret = 0;
2165         u64 offset = 0;
2166         u64 end_byte;
2167
2168         while(1) {
2169                 cache = btrfs_lookup_block_group(fs_root->fs_info, offset);
2170                 if (!cache)
2171                         break;
2172
2173                 end_byte = cache->key.objectid + cache->key.offset;
2174                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2175                         ret = relocate_extents_range(fs_root, image_root,
2176                                                      cache->key.objectid,
2177                                                      end_byte);
2178                         if (ret)
2179                                 goto fail;
2180                 }
2181                 offset = end_byte;
2182         }
2183         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2184                 offset = btrfs_sb_offset(i);
2185                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
2186
2187                 ret = relocate_extents_range(fs_root, image_root,
2188                                              offset, offset + BTRFS_STRIPE_LEN);
2189                 if (ret)
2190                         goto fail;
2191         }
2192         ret = 0;
2193 fail:
2194         return ret;
2195 }
2196
2197 static int fixup_chunk_mapping(struct btrfs_root *root)
2198 {
2199         struct btrfs_trans_handle *trans;
2200         struct btrfs_fs_info *info = root->fs_info;
2201         struct btrfs_root *chunk_root = info->chunk_root;
2202         struct extent_buffer *leaf;
2203         struct btrfs_key key;
2204         struct btrfs_path path;
2205         struct btrfs_chunk chunk;
2206         unsigned long ptr;
2207         u32 size;
2208         u64 type;
2209         int ret;
2210
2211         btrfs_init_path(&path);
2212
2213         trans = btrfs_start_transaction(root, 1);
2214         BUG_ON(!trans);
2215
2216         /*
2217          * recow the whole chunk tree. this will move all chunk tree blocks
2218          * into system block group.
2219          */
2220         memset(&key, 0, sizeof(key));
2221         while (1) {
2222                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2223                 if (ret < 0)
2224                         goto err;
2225
2226                 ret = btrfs_next_leaf(chunk_root, &path);
2227                 if (ret < 0)
2228                         goto err;
2229                 if (ret > 0)
2230                         break;
2231
2232                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2233                 btrfs_release_path(&path);
2234         }
2235         btrfs_release_path(&path);
2236
2237         /* fixup the system chunk array in super block */
2238         btrfs_set_super_sys_array_size(info->super_copy, 0);
2239
2240         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2241         key.offset = 0;
2242         key.type = BTRFS_CHUNK_ITEM_KEY;
2243
2244         ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 0);
2245         if (ret < 0)
2246                 goto err;
2247         BUG_ON(ret != 0);
2248         while(1) {
2249                 leaf = path.nodes[0];
2250                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2251                         ret = btrfs_next_leaf(chunk_root, &path);
2252                         if (ret < 0)
2253                                 goto err;
2254                         if (ret > 0)
2255                                 break;
2256                         leaf = path.nodes[0];
2257                 }
2258                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2259                 if (key.type != BTRFS_CHUNK_ITEM_KEY)
2260                         goto next;
2261
2262                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2263                 size = btrfs_item_size_nr(leaf, path.slots[0]);
2264                 BUG_ON(size != sizeof(chunk));
2265                 read_extent_buffer(leaf, &chunk, ptr, size);
2266                 type = btrfs_stack_chunk_type(&chunk);
2267
2268                 if (!(type & BTRFS_BLOCK_GROUP_SYSTEM))
2269                         goto next;
2270
2271                 ret = btrfs_add_system_chunk(trans, chunk_root, &key,
2272                                              &chunk, size);
2273                 if (ret)
2274                         goto err;
2275 next:
2276                 path.slots[0]++;
2277         }
2278
2279         ret = btrfs_commit_transaction(trans, root);
2280         BUG_ON(ret);
2281 err:
2282         btrfs_release_path(&path);
2283         return ret;
2284 }
2285
2286 static int do_convert(const char *devname, int datacsum, int packing, int noxattr,
2287                 u32 nodesize, int copylabel, const char *fslabel, int progress,
2288                 u64 features)
2289 {
2290         int i, ret, blocks_per_node;
2291         int fd = -1;
2292         u32 blocksize;
2293         u64 blocks[7];
2294         u64 total_bytes;
2295         u64 super_bytenr;
2296         ext2_filsys ext2_fs;
2297         struct btrfs_root *root;
2298         struct btrfs_root *image_root;
2299         struct task_ctx ctx;
2300         char features_buf[64];
2301         struct btrfs_mkfs_config mkfs_cfg;
2302
2303         ret = open_ext2fs(devname, &ext2_fs);
2304         if (ret) {
2305                 fprintf(stderr, "unable to open the Ext2fs\n");
2306                 goto fail;
2307         }
2308         blocksize = ext2_fs->blocksize;
2309         total_bytes = (u64)ext2_fs->super->s_blocks_count * blocksize;
2310         if (blocksize < 4096) {
2311                 fprintf(stderr, "block size is too small\n");
2312                 goto fail;
2313         }
2314         if (!(ext2_fs->super->s_feature_incompat &
2315               EXT2_FEATURE_INCOMPAT_FILETYPE)) {
2316                 fprintf(stderr, "filetype feature is missing\n");
2317                 goto fail;
2318         }
2319         if (btrfs_check_nodesize(nodesize, blocksize))
2320                 goto fail;
2321         blocks_per_node = nodesize / blocksize;
2322         ret = -blocks_per_node;
2323         for (i = 0; i < 7; i++) {
2324                 if (nodesize == blocksize)
2325                         ret = ext2_alloc_block(ext2_fs, 0, blocks + i);
2326                 else
2327                         ret = ext2_alloc_block_range(ext2_fs,
2328                                         ret + blocks_per_node, blocks_per_node,
2329                                         blocks + i);
2330                 if (ret) {
2331                         fprintf(stderr, "not enough free space\n");
2332                         goto fail;
2333                 }
2334                 blocks[i] *= blocksize;
2335         }
2336         super_bytenr = blocks[0];
2337         fd = open(devname, O_RDWR);
2338         if (fd < 0) {
2339                 fprintf(stderr, "unable to open %s\n", devname);
2340                 goto fail;
2341         }
2342         btrfs_parse_features_to_string(features_buf, features);
2343         if (features == BTRFS_MKFS_DEFAULT_FEATURES)
2344                 strcat(features_buf, " (default)");
2345
2346         printf("create btrfs filesystem:\n");
2347         printf("\tblocksize: %u\n", blocksize);
2348         printf("\tnodesize:  %u\n", nodesize);
2349         printf("\tfeatures:  %s\n", features_buf);
2350
2351         mkfs_cfg.label = ext2_fs->super->s_volume_name;
2352         mkfs_cfg.fs_uuid = NULL;
2353         memcpy(mkfs_cfg.blocks, blocks, sizeof(blocks));
2354         mkfs_cfg.num_bytes = total_bytes;
2355         mkfs_cfg.nodesize = nodesize;
2356         mkfs_cfg.sectorsize = blocksize;
2357         mkfs_cfg.stripesize = blocksize;
2358         mkfs_cfg.features = features;
2359
2360         ret = make_btrfs(fd, &mkfs_cfg);
2361         if (ret) {
2362                 fprintf(stderr, "unable to create initial ctree: %s\n",
2363                         strerror(-ret));
2364                 goto fail;
2365         }
2366         /* create a system chunk that maps the whole device */
2367         ret = prepare_system_chunk(fd, super_bytenr);
2368         if (ret) {
2369                 fprintf(stderr, "unable to update system chunk\n");
2370                 goto fail;
2371         }
2372         root = open_ctree_fd(fd, devname, super_bytenr, OPEN_CTREE_WRITES);
2373         if (!root) {
2374                 fprintf(stderr, "unable to open ctree\n");
2375                 goto fail;
2376         }
2377         ret = cache_free_extents(root, ext2_fs);
2378         if (ret) {
2379                 fprintf(stderr, "error during cache_free_extents %d\n", ret);
2380                 goto fail;
2381         }
2382         root->fs_info->extent_ops = &extent_ops;
2383         /* recover block allocation bitmap */
2384         for (i = 0; i < 7; i++) {
2385                 blocks[i] /= blocksize;
2386                 if (nodesize == blocksize)
2387                         ext2_free_block(ext2_fs, blocks[i]);
2388                 else
2389                         ext2_free_block_range(ext2_fs, blocks[i],
2390                                         blocks_per_node);
2391         }
2392         ret = init_btrfs(root);
2393         if (ret) {
2394                 fprintf(stderr, "unable to setup the root tree\n");
2395                 goto fail;
2396         }
2397         printf("creating btrfs metadata.\n");
2398         ctx.max_copy_inodes = (ext2_fs->super->s_inodes_count
2399                         - ext2_fs->super->s_free_inodes_count);
2400         ctx.cur_copy_inodes = 0;
2401
2402         if (progress) {
2403                 ctx.info = task_init(print_copied_inodes, after_copied_inodes, &ctx);
2404                 task_start(ctx.info);
2405         }
2406         ret = copy_inodes(root, ext2_fs, datacsum, packing, noxattr, &ctx);
2407         if (ret) {
2408                 fprintf(stderr, "error during copy_inodes %d\n", ret);
2409                 goto fail;
2410         }
2411         if (progress) {
2412                 task_stop(ctx.info);
2413                 task_deinit(ctx.info);
2414         }
2415         printf("creating ext2fs image file.\n");
2416         image_root = link_subvol(root, "ext2_saved", CONV_IMAGE_SUBVOL_OBJECTID);
2417         if (!image_root) {
2418                 fprintf(stderr, "unable to create subvol\n");
2419                 goto fail;
2420         }
2421         ret = create_ext2_image(image_root, ext2_fs, "image", datacsum);
2422         if (ret) {
2423                 fprintf(stderr, "error during create_ext2_image %d\n", ret);
2424                 goto fail;
2425         }
2426         memset(root->fs_info->super_copy->label, 0, BTRFS_LABEL_SIZE);
2427         if (copylabel == 1) {
2428                 strncpy(root->fs_info->super_copy->label,
2429                                 ext2_fs->super->s_volume_name, 16);
2430                 fprintf(stderr, "copy label '%s'\n",
2431                                 root->fs_info->super_copy->label);
2432         } else if (copylabel == -1) {
2433                 strncpy(root->fs_info->super_copy->label, fslabel, BTRFS_LABEL_SIZE);
2434                 fprintf(stderr, "set label to '%s'\n", fslabel);
2435         }
2436
2437         printf("cleaning up system chunk.\n");
2438         ret = cleanup_sys_chunk(root, image_root);
2439         if (ret) {
2440                 fprintf(stderr, "error during cleanup_sys_chunk %d\n", ret);
2441                 goto fail;
2442         }
2443         ret = close_ctree(root);
2444         if (ret) {
2445                 fprintf(stderr, "error during close_ctree %d\n", ret);
2446                 goto fail;
2447         }
2448         close_ext2fs(ext2_fs);
2449
2450         /*
2451          * If this step succeed, we get a mountable btrfs. Otherwise
2452          * the ext2fs is left unchanged.
2453          */
2454         ret = migrate_super_block(fd, super_bytenr, blocksize);
2455         if (ret) {
2456                 fprintf(stderr, "unable to migrate super block\n");
2457                 goto fail;
2458         }
2459
2460         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2461         if (!root) {
2462                 fprintf(stderr, "unable to open ctree\n");
2463                 goto fail;
2464         }
2465         /* move chunk tree into system chunk. */
2466         ret = fixup_chunk_mapping(root);
2467         if (ret) {
2468                 fprintf(stderr, "error during fixup_chunk_tree\n");
2469                 goto fail;
2470         }
2471         ret = close_ctree(root);
2472         close(fd);
2473
2474         printf("conversion complete.\n");
2475         return 0;
2476 fail:
2477         if (fd != -1)
2478                 close(fd);
2479         fprintf(stderr, "conversion aborted.\n");
2480         return -1;
2481 }
2482
2483 static int may_rollback(struct btrfs_root *root)
2484 {
2485         struct btrfs_fs_info *info = root->fs_info;
2486         struct btrfs_multi_bio *multi = NULL;
2487         u64 bytenr;
2488         u64 length;
2489         u64 physical;
2490         u64 total_bytes;
2491         int num_stripes;
2492         int ret;
2493
2494         if (btrfs_super_num_devices(info->super_copy) != 1)
2495                 goto fail;
2496
2497         bytenr = BTRFS_SUPER_INFO_OFFSET;
2498         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
2499
2500         while (1) {
2501                 ret = btrfs_map_block(&info->mapping_tree, WRITE, bytenr,
2502                                       &length, &multi, 0, NULL);
2503                 if (ret) {
2504                         if (ret == -ENOENT) {
2505                                 /* removed block group at the tail */
2506                                 if (length == (u64)-1)
2507                                         break;
2508
2509                                 /* removed block group in the middle */
2510                                 goto next;
2511                         }
2512                         goto fail;
2513                 }
2514
2515                 num_stripes = multi->num_stripes;
2516                 physical = multi->stripes[0].physical;
2517                 kfree(multi);
2518
2519                 if (num_stripes != 1 || physical != bytenr)
2520                         goto fail;
2521 next:
2522                 bytenr += length;
2523                 if (bytenr >= total_bytes)
2524                         break;
2525         }
2526         return 0;
2527 fail:
2528         return -1;
2529 }
2530
2531 static int do_rollback(const char *devname)
2532 {
2533         int fd = -1;
2534         int ret;
2535         int i;
2536         struct btrfs_root *root;
2537         struct btrfs_root *image_root;
2538         struct btrfs_root *chunk_root;
2539         struct btrfs_dir_item *dir;
2540         struct btrfs_inode_item *inode;
2541         struct btrfs_file_extent_item *fi;
2542         struct btrfs_trans_handle *trans;
2543         struct extent_buffer *leaf;
2544         struct btrfs_block_group_cache *cache1;
2545         struct btrfs_block_group_cache *cache2;
2546         struct btrfs_key key;
2547         struct btrfs_path path;
2548         struct extent_io_tree io_tree;
2549         char *buf = NULL;
2550         char *name;
2551         u64 bytenr;
2552         u64 num_bytes;
2553         u64 root_dir;
2554         u64 objectid;
2555         u64 offset;
2556         u64 start;
2557         u64 end;
2558         u64 sb_bytenr;
2559         u64 first_free;
2560         u64 total_bytes;
2561         u32 sectorsize;
2562
2563         extent_io_tree_init(&io_tree);
2564
2565         fd = open(devname, O_RDWR);
2566         if (fd < 0) {
2567                 fprintf(stderr, "unable to open %s\n", devname);
2568                 goto fail;
2569         }
2570         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2571         if (!root) {
2572                 fprintf(stderr, "unable to open ctree\n");
2573                 goto fail;
2574         }
2575         ret = may_rollback(root);
2576         if (ret < 0) {
2577                 fprintf(stderr, "unable to do rollback\n");
2578                 goto fail;
2579         }
2580
2581         sectorsize = root->sectorsize;
2582         buf = malloc(sectorsize);
2583         if (!buf) {
2584                 fprintf(stderr, "unable to allocate memory\n");
2585                 goto fail;
2586         }
2587
2588         btrfs_init_path(&path);
2589
2590         key.objectid = CONV_IMAGE_SUBVOL_OBJECTID;
2591         key.type = BTRFS_ROOT_ITEM_KEY;
2592         key.offset = (u64)-1;
2593         image_root = btrfs_read_fs_root(root->fs_info, &key);
2594         if (!image_root || IS_ERR(image_root)) {
2595                 fprintf(stderr, "unable to open subvol %llu\n",
2596                         (unsigned long long)key.objectid);
2597                 goto fail;
2598         }
2599
2600         name = "image";
2601         root_dir = btrfs_root_dirid(&root->root_item);
2602         dir = btrfs_lookup_dir_item(NULL, image_root, &path,
2603                                    root_dir, name, strlen(name), 0);
2604         if (!dir || IS_ERR(dir)) {
2605                 fprintf(stderr, "unable to find file %s\n", name);
2606                 goto fail;
2607         }
2608         leaf = path.nodes[0];
2609         btrfs_dir_item_key_to_cpu(leaf, dir, &key);
2610         btrfs_release_path(&path);
2611
2612         objectid = key.objectid;
2613
2614         ret = btrfs_lookup_inode(NULL, image_root, &path, &key, 0);
2615         if (ret) {
2616                 fprintf(stderr, "unable to find inode item\n");
2617                 goto fail;
2618         }
2619         leaf = path.nodes[0];
2620         inode = btrfs_item_ptr(leaf, path.slots[0], struct btrfs_inode_item);
2621         total_bytes = btrfs_inode_size(leaf, inode);
2622         btrfs_release_path(&path);
2623
2624         key.objectid = objectid;
2625         key.offset = 0;
2626         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
2627         ret = btrfs_search_slot(NULL, image_root, &key, &path, 0, 0);
2628         if (ret != 0) {
2629                 fprintf(stderr, "unable to find first file extent\n");
2630                 btrfs_release_path(&path);
2631                 goto fail;
2632         }
2633
2634         /* build mapping tree for the relocated blocks */
2635         for (offset = 0; offset < total_bytes; ) {
2636                 leaf = path.nodes[0];
2637                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2638                         ret = btrfs_next_leaf(root, &path);
2639                         if (ret != 0)
2640                                 break;
2641                         continue;
2642                 }
2643
2644                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2645                 if (key.objectid != objectid || key.offset != offset ||
2646                     btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2647                         break;
2648
2649                 fi = btrfs_item_ptr(leaf, path.slots[0],
2650                                     struct btrfs_file_extent_item);
2651                 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2652                         break;
2653                 if (btrfs_file_extent_compression(leaf, fi) ||
2654                     btrfs_file_extent_encryption(leaf, fi) ||
2655                     btrfs_file_extent_other_encoding(leaf, fi))
2656                         break;
2657
2658                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2659                 /* skip holes and direct mapped extents */
2660                 if (bytenr == 0 || bytenr == offset)
2661                         goto next_extent;
2662
2663                 bytenr += btrfs_file_extent_offset(leaf, fi);
2664                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
2665
2666                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2667                 cache2 =  btrfs_lookup_block_group(root->fs_info,
2668                                                    offset + num_bytes - 1);
2669                 if (!cache1 || cache1 != cache2 ||
2670                     (!(cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
2671                      !intersect_with_sb(offset, num_bytes)))
2672                         break;
2673
2674                 set_extent_bits(&io_tree, offset, offset + num_bytes - 1,
2675                                 EXTENT_LOCKED, GFP_NOFS);
2676                 set_state_private(&io_tree, offset, bytenr);
2677 next_extent:
2678                 offset += btrfs_file_extent_num_bytes(leaf, fi);
2679                 path.slots[0]++;
2680         }
2681         btrfs_release_path(&path);
2682
2683         if (offset < total_bytes) {
2684                 fprintf(stderr, "unable to build extent mapping\n");
2685                 goto fail;
2686         }
2687
2688         first_free = BTRFS_SUPER_INFO_OFFSET + 2 * sectorsize - 1;
2689         first_free &= ~((u64)sectorsize - 1);
2690         /* backup for extent #0 should exist */
2691         if(!test_range_bit(&io_tree, 0, first_free - 1, EXTENT_LOCKED, 1)) {
2692                 fprintf(stderr, "no backup for the first extent\n");
2693                 goto fail;
2694         }
2695         /* force no allocation from system block group */
2696         root->fs_info->system_allocs = -1;
2697         trans = btrfs_start_transaction(root, 1);
2698         BUG_ON(!trans);
2699         /*
2700          * recow the whole chunk tree, this will remove all chunk tree blocks
2701          * from system block group
2702          */
2703         chunk_root = root->fs_info->chunk_root;
2704         memset(&key, 0, sizeof(key));
2705         while (1) {
2706                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2707                 if (ret < 0)
2708                         break;
2709
2710                 ret = btrfs_next_leaf(chunk_root, &path);
2711                 if (ret)
2712                         break;
2713
2714                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2715                 btrfs_release_path(&path);
2716         }
2717         btrfs_release_path(&path);
2718
2719         offset = 0;
2720         num_bytes = 0;
2721         while(1) {
2722                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2723                 if (!cache1)
2724                         break;
2725
2726                 if (cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2727                         num_bytes += btrfs_block_group_used(&cache1->item);
2728
2729                 offset = cache1->key.objectid + cache1->key.offset;
2730         }
2731         /* only extent #0 left in system block group? */
2732         if (num_bytes > first_free) {
2733                 fprintf(stderr, "unable to empty system block group\n");
2734                 goto fail;
2735         }
2736         /* create a system chunk that maps the whole device */
2737         ret = prepare_system_chunk_sb(root->fs_info->super_copy);
2738         if (ret) {
2739                 fprintf(stderr, "unable to update system chunk\n");
2740                 goto fail;
2741         }
2742
2743         ret = btrfs_commit_transaction(trans, root);
2744         BUG_ON(ret);
2745
2746         ret = close_ctree(root);
2747         if (ret) {
2748                 fprintf(stderr, "error during close_ctree %d\n", ret);
2749                 goto fail;
2750         }
2751
2752         /* zero btrfs super block mirrors */
2753         memset(buf, 0, sectorsize);
2754         for (i = 1 ; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2755                 bytenr = btrfs_sb_offset(i);
2756                 if (bytenr >= total_bytes)
2757                         break;
2758                 ret = pwrite(fd, buf, sectorsize, bytenr);
2759                 if (ret != sectorsize) {
2760                         fprintf(stderr,
2761                                 "error during zeroing supreblock %d: %d\n",
2762                                 i, ret);
2763                         goto fail;
2764                 }
2765         }
2766
2767         sb_bytenr = (u64)-1;
2768         /* copy all relocated blocks back */
2769         while(1) {
2770                 ret = find_first_extent_bit(&io_tree, 0, &start, &end,
2771                                             EXTENT_LOCKED);
2772                 if (ret)
2773                         break;
2774
2775                 ret = get_state_private(&io_tree, start, &bytenr);
2776                 BUG_ON(ret);
2777
2778                 clear_extent_bits(&io_tree, start, end, EXTENT_LOCKED,
2779                                   GFP_NOFS);
2780
2781                 while (start <= end) {
2782                         if (start == BTRFS_SUPER_INFO_OFFSET) {
2783                                 sb_bytenr = bytenr;
2784                                 goto next_sector;
2785                         }
2786                         ret = pread(fd, buf, sectorsize, bytenr);
2787                         if (ret < 0) {
2788                                 fprintf(stderr, "error during pread %d\n", ret);
2789                                 goto fail;
2790                         }
2791                         BUG_ON(ret != sectorsize);
2792                         ret = pwrite(fd, buf, sectorsize, start);
2793                         if (ret < 0) {
2794                                 fprintf(stderr, "error during pwrite %d\n", ret);
2795                                 goto fail;
2796                         }
2797                         BUG_ON(ret != sectorsize);
2798 next_sector:
2799                         start += sectorsize;
2800                         bytenr += sectorsize;
2801                 }
2802         }
2803
2804         ret = fsync(fd);
2805         if (ret) {
2806                 fprintf(stderr, "error during fsync %d\n", ret);
2807                 goto fail;
2808         }
2809         /*
2810          * finally, overwrite btrfs super block.
2811          */
2812         ret = pread(fd, buf, sectorsize, sb_bytenr);
2813         if (ret < 0) {
2814                 fprintf(stderr, "error during pread %d\n", ret);
2815                 goto fail;
2816         }
2817         BUG_ON(ret != sectorsize);
2818         ret = pwrite(fd, buf, sectorsize, BTRFS_SUPER_INFO_OFFSET);
2819         if (ret < 0) {
2820                 fprintf(stderr, "error during pwrite %d\n", ret);
2821                 goto fail;
2822         }
2823         BUG_ON(ret != sectorsize);
2824         ret = fsync(fd);
2825         if (ret) {
2826                 fprintf(stderr, "error during fsync %d\n", ret);
2827                 goto fail;
2828         }
2829
2830         close(fd);
2831         free(buf);
2832         extent_io_tree_cleanup(&io_tree);
2833         printf("rollback complete.\n");
2834         return 0;
2835
2836 fail:
2837         if (fd != -1)
2838                 close(fd);
2839         free(buf);
2840         fprintf(stderr, "rollback aborted.\n");
2841         return -1;
2842 }
2843
2844 static void print_usage(void)
2845 {
2846         printf("usage: btrfs-convert [options] device\n");
2847         printf("options:\n");
2848         printf("\t-d|--no-datasum        disable data checksum, sets NODATASUM\n");
2849         printf("\t-i|--no-xattr          ignore xattrs and ACLs\n");
2850         printf("\t-n|--no-inline         disable inlining of small files to metadata\n");
2851         printf("\t-N|--nodesize SIZE     set filesystem metadata nodesize\n");
2852         printf("\t-r|--rollback          roll back to ext2fs\n");
2853         printf("\t-l|--label LABEL       set filesystem label\n");
2854         printf("\t-L|--copy-label        use label from converted filesystem\n");
2855         printf("\t-p|--progress          show converting progress (default)\n");
2856         printf("\t-O|--features LIST     comma separated list of filesystem features\n");
2857         printf("\t--no-progress          show only overview, not the detailed progress\n");
2858 }
2859
2860 int main(int argc, char *argv[])
2861 {
2862         int ret;
2863         int packing = 1;
2864         int noxattr = 0;
2865         int datacsum = 1;
2866         u32 nodesize = max_t(u32, sysconf(_SC_PAGESIZE),
2867                         BTRFS_MKFS_DEFAULT_NODE_SIZE);
2868         int rollback = 0;
2869         int copylabel = 0;
2870         int usage_error = 0;
2871         int progress = 1;
2872         char *file;
2873         char *fslabel = NULL;
2874         u64 features = BTRFS_MKFS_DEFAULT_FEATURES;
2875
2876         while(1) {
2877                 enum { GETOPT_VAL_NO_PROGRESS = 256 };
2878                 static const struct option long_options[] = {
2879                         { "no-progress", no_argument, NULL,
2880                                 GETOPT_VAL_NO_PROGRESS },
2881                         { "no-datasum", no_argument, NULL, 'd' },
2882                         { "no-inline", no_argument, NULL, 'n' },
2883                         { "no-xattr", no_argument, NULL, 'i' },
2884                         { "rollback", no_argument, NULL, 'r' },
2885                         { "features", required_argument, NULL, 'O' },
2886                         { "progress", no_argument, NULL, 'p' },
2887                         { "label", required_argument, NULL, 'l' },
2888                         { "copy-label", no_argument, NULL, 'L' },
2889                         { "nodesize", required_argument, NULL, 'N' },
2890                         { "help", no_argument, NULL, GETOPT_VAL_HELP},
2891                         { NULL, 0, NULL, 0 }
2892                 };
2893                 int c = getopt_long(argc, argv, "dinN:rl:LpO:", long_options, NULL);
2894
2895                 if (c < 0)
2896                         break;
2897                 switch(c) {
2898                         case 'd':
2899                                 datacsum = 0;
2900                                 break;
2901                         case 'i':
2902                                 noxattr = 1;
2903                                 break;
2904                         case 'n':
2905                                 packing = 0;
2906                                 break;
2907                         case 'N':
2908                                 nodesize = parse_size(optarg);
2909                                 break;
2910                         case 'r':
2911                                 rollback = 1;
2912                                 break;
2913                         case 'l':
2914                                 copylabel = -1;
2915                                 fslabel = strdup(optarg);
2916                                 if (strlen(fslabel) > BTRFS_LABEL_SIZE) {
2917                                         fprintf(stderr,
2918                                                 "warning: label too long, trimmed to %d bytes\n",
2919                                                 BTRFS_LABEL_SIZE);
2920                                         fslabel[BTRFS_LABEL_SIZE] = 0;
2921                                 }
2922                                 break;
2923                         case 'L':
2924                                 copylabel = 1;
2925                                 break;
2926                         case 'p':
2927                                 progress = 1;
2928                                 break;
2929                         case 'O': {
2930                                 char *orig = strdup(optarg);
2931                                 char *tmp = orig;
2932
2933                                 tmp = btrfs_parse_fs_features(tmp, &features);
2934                                 if (tmp) {
2935                                         fprintf(stderr,
2936                                                 "Unrecognized filesystem feature '%s'\n",
2937                                                         tmp);
2938                                         free(orig);
2939                                         exit(1);
2940                                 }
2941                                 free(orig);
2942                                 if (features & BTRFS_FEATURE_LIST_ALL) {
2943                                         btrfs_list_all_fs_features(
2944                                                 ~BTRFS_CONVERT_ALLOWED_FEATURES);
2945                                         exit(0);
2946                                 }
2947                                 if (features & ~BTRFS_CONVERT_ALLOWED_FEATURES) {
2948                                         char buf[64];
2949
2950                                         btrfs_parse_features_to_string(buf,
2951                                                 features & ~BTRFS_CONVERT_ALLOWED_FEATURES);
2952                                         fprintf(stderr,
2953                                                 "ERROR: features not allowed for convert: %s\n",
2954                                                 buf);
2955                                         exit(1);
2956                                 }
2957
2958                                 break;
2959                                 }
2960                         case GETOPT_VAL_NO_PROGRESS:
2961                                 progress = 0;
2962                                 break;
2963                         case GETOPT_VAL_HELP:
2964                         default:
2965                                 print_usage();
2966                                 return c != GETOPT_VAL_HELP;
2967                 }
2968         }
2969         argc = argc - optind;
2970         set_argv0(argv);
2971         if (check_argc_exact(argc, 1)) {
2972                 print_usage();
2973                 return 1;
2974         }
2975
2976         if (rollback && (!datacsum || noxattr || !packing)) {
2977                 fprintf(stderr,
2978                         "Usage error: -d, -i, -n options do not apply to rollback\n");
2979                 usage_error++;
2980         }
2981
2982         if (usage_error) {
2983                 print_usage();
2984                 return 1;
2985         }
2986
2987         file = argv[optind];
2988         ret = check_mounted(file);
2989         if (ret < 0) {
2990                 fprintf(stderr, "Could not check mount status: %s\n",
2991                         strerror(-ret));
2992                 return 1;
2993         } else if (ret) {
2994                 fprintf(stderr, "%s is mounted\n", file);
2995                 return 1;
2996         }
2997
2998         if (rollback) {
2999                 ret = do_rollback(file);
3000         } else {
3001                 ret = do_convert(file, datacsum, packing, noxattr, nodesize,
3002                                 copylabel, fslabel, progress, features);
3003         }
3004         if (ret)
3005                 return 1;
3006         return 0;
3007 }