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 EXT2_IMAGE_SUBVOL_OBJECTID BTRFS_FIRST_FREE_OBJECTID
  46
  47 struct task_ctx {
  48         uint32_t max_copy_inodes;
  49         uint32_t cur_copy_inodes;
  50         struct task_info *info;
  51 };
  52
  53 static void *print_copied_inodes(void *p)
  54 {
  55         struct task_ctx *priv = p;
  56         const char work_indicator[] = { '.', 'o', 'O', 'o' };
  57         uint32_t count = 0;
  58
  59         task_period_start(priv->info, 1000 /* 1s */);
  60         while (1) {
  61                 count++;
  62                 printf("copy inodes [%c] [%10d/%10d]\r",
  63                        work_indicator[count % 4], priv->cur_copy_inodes,
  64                        priv->max_copy_inodes);
  65                 fflush(stdout);
  66                 task_period_wait(priv->info);
  67         }
  68
  69         return NULL;
  70 }
  71
  72 static int after_copied_inodes(void *p)
  73 {
  74         struct task_ctx *priv = p;
  75
  76         printf("\n");
  77         task_period_stop(priv->info);
  78
  79         return 0;
  80 }
  81
  82 /*
  83  * Open Ext2fs in readonly mode, read block allocation bitmap and
  84  * inode bitmap into memory.
  85  */
  86 static int open_ext2fs(const char *name, ext2_filsys *ret_fs)
  87 {
  88         errcode_t ret;
  89         ext2_filsys ext2_fs;
  90         ext2_ino_t ino;
  91         ret = ext2fs_open(name, 0, 0, 0, unix_io_manager, &ext2_fs);
  92         if (ret) {
  93                 fprintf(stderr, "ext2fs_open: %s\n", error_message(ret));
  94                 goto fail;
  95         }
  96         ret = ext2fs_read_inode_bitmap(ext2_fs);
  97         if (ret) {
  98                 fprintf(stderr, "ext2fs_read_inode_bitmap: %s\n",
  99                         error_message(ret));
 100                 goto fail;
 101         }
 102         ret = ext2fs_read_block_bitmap(ext2_fs);
 103         if (ret) {
 104                 fprintf(stderr, "ext2fs_read_block_bitmap: %s\n",
 105                         error_message(ret));
 106                 goto fail;
 107         }
 108         /*
 109          * search each block group for a free inode. this set up
 110          * uninit block/inode bitmaps appropriately.
 111          */
 112         ino = 1;
 113         while (ino <= ext2_fs->super->s_inodes_count) {
 114                 ext2_ino_t foo;
 115                 ext2fs_new_inode(ext2_fs, ino, 0, NULL, &foo);
 116                 ino += EXT2_INODES_PER_GROUP(ext2_fs->super);
 117         }
 118
 119         *ret_fs = ext2_fs;
 120         return 0;
 121 fail:
 122         return -1;
 123 }
 124
 125 static int close_ext2fs(ext2_filsys fs)
 126 {
 127         ext2fs_close(fs);
 128         return 0;
 129 }
 130
 131 static int ext2_alloc_block(ext2_filsys fs, u64 goal, u64 *block_ret)
 132 {
 133         blk_t block;
 134
 135         if (!ext2fs_new_block(fs, goal, NULL, &block)) {
 136                 ext2fs_fast_mark_block_bitmap(fs->block_map, block);
 137                 *block_ret = block;
 138                 return 0;
 139         }
 140         return -ENOSPC;
 141 }
 142
 143 static int ext2_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 {
 212         u64 start;
 213         u64 end;
 214         u64 last = hint_byte;
 215         int ret;
 216         int wrapped = 0;
 217         struct btrfs_block_group_cache *cache;
 218
 219         while(1) {
 220                 ret = find_first_extent_bit(&root->fs_info->free_space_cache,
 221                                             last, &start, &end, EXTENT_DIRTY);
 222                 if (ret) {
 223                         if (wrapped++ == 0) {
 224                                 last = 0;
 225                                 continue;
 226                         } else {
 227                                 goto fail;
 228                         }
 229                 }
 230
 231                 start = max(last, start);
 232                 last = end + 1;
 233                 if (last - start < num_bytes)
 234                         continue;
 235
 236                 last = start + num_bytes;
 237                 if (test_range_bit(&root->fs_info->pinned_extents,
 238                                    start, last - 1, EXTENT_DIRTY, 0))
 239                         continue;
 240
 241                 cache = btrfs_lookup_block_group(root->fs_info, start);
 242                 BUG_ON(!cache);
 243                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM ||
 244                     last > cache->key.objectid + cache->key.offset) {
 245                         last = cache->key.objectid + cache->key.offset;
 246                         continue;
 247                 }
 248
 249                 clear_extent_dirty(&root->fs_info->free_space_cache,
 250                                    start, start + num_bytes - 1, 0);
 251
 252                 ins->objectid = start;
 253                 ins->offset = num_bytes;
 254                 ins->type = BTRFS_EXTENT_ITEM_KEY;
 255                 return 0;
 256         }
 257 fail:
 258         fprintf(stderr, "not enough free space\n");
 259         return -ENOSPC;
 260 }
 261
 262 static int intersect_with_sb(u64 bytenr, u64 num_bytes)
 263 {
 264         int i;
 265         u64 offset;
 266
 267         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 268                 offset = btrfs_sb_offset(i);
 269                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
 270
 271                 if (bytenr < offset + BTRFS_STRIPE_LEN &&
 272                     bytenr + num_bytes > offset)
 273                         return 1;
 274         }
 275         return 0;
 276 }
 277
 278 static int custom_free_extent(struct btrfs_root *root, u64 bytenr,
 279                               u64 num_bytes)
 280 {
 281         return intersect_with_sb(bytenr, num_bytes);
 282 }
 283
 284 static struct btrfs_extent_ops extent_ops = {
 285         .alloc_extent = custom_alloc_extent,
 286         .free_extent = custom_free_extent,
 287 };
 288
 289 struct dir_iterate_data {
 290         struct btrfs_trans_handle *trans;
 291         struct btrfs_root *root;
 292         struct btrfs_inode_item *inode;
 293         u64 objectid;
 294         u64 index_cnt;
 295         u64 parent;
 296         int errcode;
 297 };
 298
 299 static u8 filetype_conversion_table[EXT2_FT_MAX] = {
 300         [EXT2_FT_UNKNOWN]       = BTRFS_FT_UNKNOWN,
 301         [EXT2_FT_REG_FILE]      = BTRFS_FT_REG_FILE,
 302         [EXT2_FT_DIR]           = BTRFS_FT_DIR,
 303         [EXT2_FT_CHRDEV]        = BTRFS_FT_CHRDEV,
 304         [EXT2_FT_BLKDEV]        = BTRFS_FT_BLKDEV,
 305         [EXT2_FT_FIFO]          = BTRFS_FT_FIFO,
 306         [EXT2_FT_SOCK]          = BTRFS_FT_SOCK,
 307         [EXT2_FT_SYMLINK]       = BTRFS_FT_SYMLINK,
 308 };
 309
 310 static int dir_iterate_proc(ext2_ino_t dir, int entry,
 311                             struct ext2_dir_entry *dirent,
 312                             int offset, int blocksize,
 313                             char *buf,void *priv_data)
 314 {
 315         int ret;
 316         int file_type;
 317         u64 objectid;
 318         u64 inode_size;
 319         char dotdot[] = "..";
 320         struct btrfs_key location;
 321         struct dir_iterate_data *idata = (struct dir_iterate_data *)priv_data;
 322         int name_len;
 323
 324         name_len = dirent->name_len & 0xFF;
 325
 326         objectid = dirent->inode + INO_OFFSET;
 327         if (!strncmp(dirent->name, dotdot, name_len)) {
 328                 if (name_len == 2) {
 329                         BUG_ON(idata->parent != 0);
 330                         idata->parent = objectid;
 331                 }
 332                 return 0;
 333         }
 334         if (dirent->inode < EXT2_GOOD_OLD_FIRST_INO)
 335                 return 0;
 336
 337         location.objectid = objectid;
 338         location.offset = 0;
 339         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
 340
 341         file_type = dirent->name_len >> 8;
 342         BUG_ON(file_type > EXT2_FT_SYMLINK);
 343         ret = btrfs_insert_dir_item(idata->trans, idata->root,
 344                                     dirent->name, name_len,
 345                                     idata->objectid, &location,
 346                                     filetype_conversion_table[file_type],
 347                                     idata->index_cnt);
 348         if (ret)
 349                 goto fail;
 350         ret = btrfs_insert_inode_ref(idata->trans, idata->root,
 351                                      dirent->name, name_len,
 352                                      objectid, idata->objectid,
 353                                      idata->index_cnt);
 354         if (ret)
 355                 goto fail;
 356         idata->index_cnt++;
 357         inode_size = btrfs_stack_inode_size(idata->inode) +
 358                      name_len * 2;
 359         btrfs_set_stack_inode_size(idata->inode, inode_size);
 360         return 0;
 361 fail:
 362         idata->errcode = ret;
 363         return BLOCK_ABORT;
 364 }
 365
 366 static int create_dir_entries(struct btrfs_trans_handle *trans,
 367                               struct btrfs_root *root, u64 objectid,
 368                               struct btrfs_inode_item *btrfs_inode,
 369                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 370 {
 371         int ret;
 372         errcode_t err;
 373         struct dir_iterate_data data = {
 374                 .trans          = trans,
 375                 .root           = root,
 376                 .inode          = btrfs_inode,
 377                 .objectid       = objectid,
 378                 .index_cnt      = 2,
 379                 .parent         = 0,
 380                 .errcode        = 0,
 381         };
 382
 383         err = ext2fs_dir_iterate2(ext2_fs, ext2_ino, 0, NULL,
 384                                   dir_iterate_proc, &data);
 385         if (err)
 386                 goto error;
 387         ret = data.errcode;
 388         if (ret == 0 && data.parent == objectid) {
 389                 ret = btrfs_insert_inode_ref(trans, root, "..", 2,
 390                                              objectid, objectid, 0);
 391         }
 392         return ret;
 393 error:
 394         fprintf(stderr, "ext2fs_dir_iterate2: %s\n", error_message(err));
 395         return -1;
 396 }
 397
 398 static int read_disk_extent(struct btrfs_root *root, u64 bytenr,
 399                             u32 num_bytes, char *buffer)
 400 {
 401         int ret;
 402         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
 403
 404         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, bytenr);
 405         if (ret != num_bytes)
 406                 goto fail;
 407         ret = 0;
 408 fail:
 409         if (ret > 0)
 410                 ret = -1;
 411         return ret;
 412 }
 413
 414 static int csum_disk_extent(struct btrfs_trans_handle *trans,
 415                             struct btrfs_root *root,
 416                             u64 disk_bytenr, u64 num_bytes)
 417 {
 418         u32 blocksize = root->sectorsize;
 419         u64 offset;
 420         char *buffer;
 421         int ret = 0;
 422
 423         buffer = malloc(blocksize);
 424         if (!buffer)
 425                 return -ENOMEM;
 426         for (offset = 0; offset < num_bytes; offset += blocksize) {
 427                 ret = read_disk_extent(root, disk_bytenr + offset,
 428                                         blocksize, buffer);
 429                 if (ret)
 430                         break;
 431                 ret = btrfs_csum_file_block(trans,
 432                                             root->fs_info->csum_root,
 433                                             disk_bytenr + num_bytes,
 434                                             disk_bytenr + offset,
 435                                             buffer, blocksize);
 436                 if (ret)
 437                         break;
 438         }
 439         free(buffer);
 440         return ret;
 441 }
 442
 443 static int record_file_blocks(struct btrfs_trans_handle *trans,
 444                               struct btrfs_root *root, u64 objectid,
 445                               struct btrfs_inode_item *inode,
 446                               u64 file_block, u64 disk_block,
 447                               u64 num_blocks, int checksum)
 448 {
 449         int ret;
 450         u64 file_pos = file_block * root->sectorsize;
 451         u64 disk_bytenr = disk_block * root->sectorsize;
 452         u64 num_bytes = num_blocks * root->sectorsize;
 453         ret = btrfs_record_file_extent(trans, root, objectid, inode, file_pos,
 454                                         disk_bytenr, num_bytes);
 455
 456         if (ret || !checksum || disk_bytenr == 0)
 457                 return ret;
 458
 459         return csum_disk_extent(trans, root, disk_bytenr, num_bytes);
 460 }
 461
 462 struct blk_iterate_data {
 463         struct btrfs_trans_handle *trans;
 464         struct btrfs_root *root;
 465         struct btrfs_inode_item *inode;
 466         u64 objectid;
 467         u64 first_block;
 468         u64 disk_block;
 469         u64 num_blocks;
 470         u64 boundary;
 471         int checksum;
 472         int errcode;
 473 };
 474
 475 static int block_iterate_proc(ext2_filsys ext2_fs,
 476                               u64 disk_block, u64 file_block,
 477                               struct blk_iterate_data *idata)
 478 {
 479         int ret;
 480         int sb_region;
 481         int do_barrier;
 482         struct btrfs_root *root = idata->root;
 483         struct btrfs_trans_handle *trans = idata->trans;
 484         struct btrfs_block_group_cache *cache;
 485         u64 bytenr = disk_block * root->sectorsize;
 486
 487         sb_region = intersect_with_sb(bytenr, root->sectorsize);
 488         do_barrier = sb_region || disk_block >= idata->boundary;
 489         if ((idata->num_blocks > 0 && do_barrier) ||
 490             (file_block > idata->first_block + idata->num_blocks) ||
 491             (disk_block != idata->disk_block + idata->num_blocks)) {
 492                 if (idata->num_blocks > 0) {
 493                         ret = record_file_blocks(trans, root, idata->objectid,
 494                                         idata->inode, idata->first_block,
 495                                         idata->disk_block, idata->num_blocks,
 496                                         idata->checksum);
 497                         if (ret)
 498                                 goto fail;
 499                         idata->first_block += idata->num_blocks;
 500                         idata->num_blocks = 0;
 501                 }
 502                 if (file_block > idata->first_block) {
 503                         ret = record_file_blocks(trans, root, idata->objectid,
 504                                         idata->inode, idata->first_block,
 505                                         0, file_block - idata->first_block,
 506                                         idata->checksum);
 507                         if (ret)
 508                                 goto fail;
 509                 }
 510
 511                 if (sb_region) {
 512                         bytenr += BTRFS_STRIPE_LEN - 1;
 513                         bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 514                 } else {
 515                         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
 516                         BUG_ON(!cache);
 517                         bytenr = cache->key.objectid + cache->key.offset;
 518                 }
 519
 520                 idata->first_block = file_block;
 521                 idata->disk_block = disk_block;
 522                 idata->boundary = bytenr / root->sectorsize;
 523         }
 524         idata->num_blocks++;
 525         return 0;
 526 fail:
 527         idata->errcode = ret;
 528         return BLOCK_ABORT;
 529 }
 530
 531 static int __block_iterate_proc(ext2_filsys fs, blk_t *blocknr,
 532                                 e2_blkcnt_t blockcnt, blk_t ref_block,
 533                                 int ref_offset, void *priv_data)
 534 {
 535         struct blk_iterate_data *idata;
 536         idata = (struct blk_iterate_data *)priv_data;
 537         return block_iterate_proc(fs, *blocknr, blockcnt, idata);
 538 }
 539
 540 /*
 541  * traverse file's data blocks, record these data blocks as file extents.
 542  */
 543 static int create_file_extents(struct btrfs_trans_handle *trans,
 544                                struct btrfs_root *root, u64 objectid,
 545                                struct btrfs_inode_item *btrfs_inode,
 546                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 547                                int datacsum, int packing)
 548 {
 549         int ret;
 550         char *buffer = NULL;
 551         errcode_t err;
 552         u32 last_block;
 553         u32 sectorsize = root->sectorsize;
 554         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 555         struct blk_iterate_data data = {
 556                 .trans          = trans,
 557                 .root           = root,
 558                 .inode          = btrfs_inode,
 559                 .objectid       = objectid,
 560                 .first_block    = 0,
 561                 .disk_block     = 0,
 562                 .num_blocks     = 0,
 563                 .boundary       = (u64)-1,
 564                 .checksum       = datacsum,
 565                 .errcode        = 0,
 566         };
 567         err = ext2fs_block_iterate2(ext2_fs, ext2_ino, BLOCK_FLAG_DATA_ONLY,
 568                                     NULL, __block_iterate_proc, &data);
 569         if (err)
 570                 goto error;
 571         ret = data.errcode;
 572         if (ret)
 573                 goto fail;
 574         if (packing && data.first_block == 0 && data.num_blocks > 0 &&
 575             inode_size <= BTRFS_MAX_INLINE_DATA_SIZE(root)) {
 576                 u64 num_bytes = data.num_blocks * sectorsize;
 577                 u64 disk_bytenr = data.disk_block * sectorsize;
 578                 u64 nbytes;
 579
 580                 buffer = malloc(num_bytes);
 581                 if (!buffer)
 582                         return -ENOMEM;
 583                 ret = read_disk_extent(root, disk_bytenr, num_bytes, buffer);
 584                 if (ret)
 585                         goto fail;
 586                 if (num_bytes > inode_size)
 587                         num_bytes = inode_size;
 588                 ret = btrfs_insert_inline_extent(trans, root, objectid,
 589                                                  0, buffer, num_bytes);
 590                 if (ret)
 591                         goto fail;
 592                 nbytes = btrfs_stack_inode_nbytes(btrfs_inode) + num_bytes;
 593                 btrfs_set_stack_inode_nbytes(btrfs_inode, nbytes);
 594         } else if (data.num_blocks > 0) {
 595                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 596                                          data.first_block, data.disk_block,
 597                                          data.num_blocks, data.checksum);
 598                 if (ret)
 599                         goto fail;
 600         }
 601         data.first_block += data.num_blocks;
 602         last_block = (inode_size + sectorsize - 1) / sectorsize;
 603         if (last_block > data.first_block) {
 604                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 605                                          data.first_block, 0, last_block -
 606                                          data.first_block, data.checksum);
 607         }
 608 fail:
 609         free(buffer);
 610         return ret;
 611 error:
 612         fprintf(stderr, "ext2fs_block_iterate2: %s\n", error_message(err));
 613         return -1;
 614 }
 615
 616 static int create_symbol_link(struct btrfs_trans_handle *trans,
 617                               struct btrfs_root *root, u64 objectid,
 618                               struct btrfs_inode_item *btrfs_inode,
 619                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 620                               struct ext2_inode *ext2_inode)
 621 {
 622         int ret;
 623         char *pathname;
 624         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 625         if (ext2fs_inode_data_blocks(ext2_fs, ext2_inode)) {
 626                 btrfs_set_stack_inode_size(btrfs_inode, inode_size + 1);
 627                 ret = create_file_extents(trans, root, objectid, btrfs_inode,
 628                                           ext2_fs, ext2_ino, 1, 1);
 629                 btrfs_set_stack_inode_size(btrfs_inode, inode_size);
 630                 return ret;
 631         }
 632
 633         pathname = (char *)&(ext2_inode->i_block[0]);
 634         BUG_ON(pathname[inode_size] != 0);
 635         ret = btrfs_insert_inline_extent(trans, root, objectid, 0,
 636                                          pathname, inode_size + 1);
 637         btrfs_set_stack_inode_nbytes(btrfs_inode, inode_size + 1);
 638         return ret;
 639 }
 640
 641 /*
 642  * Following xattr/acl related codes are based on codes in
 643  * fs/ext3/xattr.c and fs/ext3/acl.c
 644  */
 645 #define EXT2_XATTR_BHDR(ptr) ((struct ext2_ext_attr_header *)(ptr))
 646 #define EXT2_XATTR_BFIRST(ptr) \
 647         ((struct ext2_ext_attr_entry *)(EXT2_XATTR_BHDR(ptr) + 1))
 648 #define EXT2_XATTR_IHDR(inode) \
 649         ((struct ext2_ext_attr_header *) ((void *)(inode) + \
 650                 EXT2_GOOD_OLD_INODE_SIZE + (inode)->i_extra_isize))
 651 #define EXT2_XATTR_IFIRST(inode) \
 652         ((struct ext2_ext_attr_entry *) ((void *)EXT2_XATTR_IHDR(inode) + \
 653                 sizeof(EXT2_XATTR_IHDR(inode)->h_magic)))
 654
 655 static int ext2_xattr_check_names(struct ext2_ext_attr_entry *entry,
 656                                   const void *end)
 657 {
 658         struct ext2_ext_attr_entry *next;
 659
 660         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 661                 next = EXT2_EXT_ATTR_NEXT(entry);
 662                 if ((void *)next >= end)
 663                         return -EIO;
 664                 entry = next;
 665         }
 666         return 0;
 667 }
 668
 669 static int ext2_xattr_check_block(const char *buf, size_t size)
 670 {
 671         int error;
 672         struct ext2_ext_attr_header *header = EXT2_XATTR_BHDR(buf);
 673
 674         if (header->h_magic != EXT2_EXT_ATTR_MAGIC ||
 675             header->h_blocks != 1)
 676                 return -EIO;
 677         error = ext2_xattr_check_names(EXT2_XATTR_BFIRST(buf), buf + size);
 678         return error;
 679 }
 680
 681 static int ext2_xattr_check_entry(struct ext2_ext_attr_entry *entry,
 682                                   size_t size)
 683 {
 684         size_t value_size = entry->e_value_size;
 685
 686         if (entry->e_value_block != 0 || value_size > size ||
 687             entry->e_value_offs + value_size > size)
 688                 return -EIO;
 689         return 0;
 690 }
 691
 692 #define EXT2_ACL_VERSION        0x0001
 693
 694 /* 23.2.5 acl_tag_t values */
 695
 696 #define ACL_UNDEFINED_TAG       (0x00)
 697 #define ACL_USER_OBJ            (0x01)
 698 #define ACL_USER                (0x02)
 699 #define ACL_GROUP_OBJ           (0x04)
 700 #define ACL_GROUP               (0x08)
 701 #define ACL_MASK                (0x10)
 702 #define ACL_OTHER               (0x20)
 703
 704 /* 23.2.7 ACL qualifier constants */
 705
 706 #define ACL_UNDEFINED_ID        ((id_t)-1)
 707
 708 typedef struct {
 709         __le16          e_tag;
 710         __le16          e_perm;
 711         __le32          e_id;
 712 } ext2_acl_entry;
 713
 714 typedef struct {
 715         __le16          e_tag;
 716         __le16          e_perm;
 717 } ext2_acl_entry_short;
 718
 719 typedef struct {
 720         __le32          a_version;
 721 } ext2_acl_header;
 722
 723 static inline int ext2_acl_count(size_t size)
 724 {
 725         ssize_t s;
 726         size -= sizeof(ext2_acl_header);
 727         s = size - 4 * sizeof(ext2_acl_entry_short);
 728         if (s < 0) {
 729                 if (size % sizeof(ext2_acl_entry_short))
 730                         return -1;
 731                 return size / sizeof(ext2_acl_entry_short);
 732         } else {
 733                 if (s % sizeof(ext2_acl_entry))
 734                         return -1;
 735                 return s / sizeof(ext2_acl_entry) + 4;
 736         }
 737 }
 738
 739 #define ACL_EA_VERSION          0x0002
 740
 741 typedef struct {
 742         __le16          e_tag;
 743         __le16          e_perm;
 744         __le32          e_id;
 745 } acl_ea_entry;
 746
 747 typedef struct {
 748         __le32          a_version;
 749         acl_ea_entry    a_entries[0];
 750 } acl_ea_header;
 751
 752 static inline size_t acl_ea_size(int count)
 753 {
 754         return sizeof(acl_ea_header) + count * sizeof(acl_ea_entry);
 755 }
 756
 757 static int ext2_acl_to_xattr(void *dst, const void *src,
 758                              size_t dst_size, size_t src_size)
 759 {
 760         int i, count;
 761         const void *end = src + src_size;
 762         acl_ea_header *ext_acl = (acl_ea_header *)dst;
 763         acl_ea_entry *dst_entry = ext_acl->a_entries;
 764         ext2_acl_entry *src_entry;
 765
 766         if (src_size < sizeof(ext2_acl_header))
 767                 goto fail;
 768         if (((ext2_acl_header *)src)->a_version !=
 769             cpu_to_le32(EXT2_ACL_VERSION))
 770                 goto fail;
 771         src += sizeof(ext2_acl_header);
 772         count = ext2_acl_count(src_size);
 773         if (count <= 0)
 774                 goto fail;
 775
 776         BUG_ON(dst_size < acl_ea_size(count));
 777         ext_acl->a_version = cpu_to_le32(ACL_EA_VERSION);
 778         for (i = 0; i < count; i++, dst_entry++) {
 779                 src_entry = (ext2_acl_entry *)src;
 780                 if (src + sizeof(ext2_acl_entry_short) > end)
 781                         goto fail;
 782                 dst_entry->e_tag = src_entry->e_tag;
 783                 dst_entry->e_perm = src_entry->e_perm;
 784                 switch (le16_to_cpu(src_entry->e_tag)) {
 785                 case ACL_USER_OBJ:
 786                 case ACL_GROUP_OBJ:
 787                 case ACL_MASK:
 788                 case ACL_OTHER:
 789                         src += sizeof(ext2_acl_entry_short);
 790                         dst_entry->e_id = cpu_to_le32(ACL_UNDEFINED_ID);
 791                         break;
 792                 case ACL_USER:
 793                 case ACL_GROUP:
 794                         src += sizeof(ext2_acl_entry);
 795                         if (src > end)
 796                                 goto fail;
 797                         dst_entry->e_id = src_entry->e_id;
 798                         break;
 799                 default:
 800                         goto fail;
 801                 }
 802         }
 803         if (src != end)
 804                 goto fail;
 805         return 0;
 806 fail:
 807         return -EINVAL;
 808 }
 809
 810 static char *xattr_prefix_table[] = {
 811         [1] =   "user.",
 812         [2] =   "system.posix_acl_access",
 813         [3] =   "system.posix_acl_default",
 814         [4] =   "trusted.",
 815         [6] =   "security.",
 816 };
 817
 818 static int copy_single_xattr(struct btrfs_trans_handle *trans,
 819                              struct btrfs_root *root, u64 objectid,
 820                              struct ext2_ext_attr_entry *entry,
 821                              const void *data, u32 datalen)
 822 {
 823         int ret = 0;
 824         int name_len;
 825         int name_index;
 826         void *databuf = NULL;
 827         char namebuf[XATTR_NAME_MAX + 1];
 828
 829         name_index = entry->e_name_index;
 830         if (name_index >= ARRAY_SIZE(xattr_prefix_table) ||
 831             xattr_prefix_table[name_index] == NULL)
 832                 return -EOPNOTSUPP;
 833         name_len = strlen(xattr_prefix_table[name_index]) +
 834                    entry->e_name_len;
 835         if (name_len >= sizeof(namebuf))
 836                 return -ERANGE;
 837
 838         if (name_index == 2 || name_index == 3) {
 839                 size_t bufsize = acl_ea_size(ext2_acl_count(datalen));
 840                 databuf = malloc(bufsize);
 841                 if (!databuf)
 842                        return -ENOMEM;
 843                 ret = ext2_acl_to_xattr(databuf, data, bufsize, datalen);
 844                 if (ret)
 845                         goto out;
 846                 data = databuf;
 847                 datalen = bufsize;
 848         }
 849         strncpy(namebuf, xattr_prefix_table[name_index], XATTR_NAME_MAX);
 850         strncat(namebuf, EXT2_EXT_ATTR_NAME(entry), entry->e_name_len);
 851         if (name_len + datalen > BTRFS_LEAF_DATA_SIZE(root) -
 852             sizeof(struct btrfs_item) - sizeof(struct btrfs_dir_item)) {
 853                 fprintf(stderr, "skip large xattr on inode %Lu name %.*s\n",
 854                         objectid - INO_OFFSET, name_len, namebuf);
 855                 goto out;
 856         }
 857         ret = btrfs_insert_xattr_item(trans, root, namebuf, name_len,
 858                                       data, datalen, objectid);
 859 out:
 860         free(databuf);
 861         return ret;
 862 }
 863
 864 static int copy_extended_attrs(struct btrfs_trans_handle *trans,
 865                                struct btrfs_root *root, u64 objectid,
 866                                struct btrfs_inode_item *btrfs_inode,
 867                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 868 {
 869         int ret = 0;
 870         int inline_ea = 0;
 871         errcode_t err;
 872         u32 datalen;
 873         u32 block_size = ext2_fs->blocksize;
 874         u32 inode_size = EXT2_INODE_SIZE(ext2_fs->super);
 875         struct ext2_inode_large *ext2_inode;
 876         struct ext2_ext_attr_entry *entry;
 877         void *data;
 878         char *buffer = NULL;
 879         char inode_buf[EXT2_GOOD_OLD_INODE_SIZE];
 880
 881         if (inode_size <= EXT2_GOOD_OLD_INODE_SIZE) {
 882                 ext2_inode = (struct ext2_inode_large *)inode_buf;
 883         } else {
 884                 ext2_inode = (struct ext2_inode_large *)malloc(inode_size);
 885                 if (!ext2_inode)
 886                        return -ENOMEM;
 887         }
 888         err = ext2fs_read_inode_full(ext2_fs, ext2_ino, (void *)ext2_inode,
 889                                      inode_size);
 890         if (err) {
 891                 fprintf(stderr, "ext2fs_read_inode_full: %s\n",
 892                         error_message(err));
 893                 ret = -1;
 894                 goto out;
 895         }
 896
 897         if (ext2_ino > ext2_fs->super->s_first_ino &&
 898             inode_size > EXT2_GOOD_OLD_INODE_SIZE) {
 899                 if (EXT2_GOOD_OLD_INODE_SIZE +
 900                     ext2_inode->i_extra_isize > inode_size) {
 901                         ret = -EIO;
 902                         goto out;
 903                 }
 904                 if (ext2_inode->i_extra_isize != 0 &&
 905                     EXT2_XATTR_IHDR(ext2_inode)->h_magic ==
 906                     EXT2_EXT_ATTR_MAGIC) {
 907                         inline_ea = 1;
 908                 }
 909         }
 910         if (inline_ea) {
 911                 int total;
 912                 void *end = (void *)ext2_inode + inode_size;
 913                 entry = EXT2_XATTR_IFIRST(ext2_inode);
 914                 total = end - (void *)entry;
 915                 ret = ext2_xattr_check_names(entry, end);
 916                 if (ret)
 917                         goto out;
 918                 while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 919                         ret = ext2_xattr_check_entry(entry, total);
 920                         if (ret)
 921                                 goto out;
 922                         data = (void *)EXT2_XATTR_IFIRST(ext2_inode) +
 923                                 entry->e_value_offs;
 924                         datalen = entry->e_value_size;
 925                         ret = copy_single_xattr(trans, root, objectid,
 926                                                 entry, data, datalen);
 927                         if (ret)
 928                                 goto out;
 929                         entry = EXT2_EXT_ATTR_NEXT(entry);
 930                 }
 931         }
 932
 933         if (ext2_inode->i_file_acl == 0)
 934                 goto out;
 935
 936         buffer = malloc(block_size);
 937         if (!buffer) {
 938                 ret = -ENOMEM;
 939                 goto out;
 940         }
 941         err = ext2fs_read_ext_attr(ext2_fs, ext2_inode->i_file_acl, buffer);
 942         if (err) {
 943                 fprintf(stderr, "ext2fs_read_ext_attr: %s\n",
 944                         error_message(err));
 945                 ret = -1;
 946                 goto out;
 947         }
 948         ret = ext2_xattr_check_block(buffer, block_size);
 949         if (ret)
 950                 goto out;
 951
 952         entry = EXT2_XATTR_BFIRST(buffer);
 953         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 954                 ret = ext2_xattr_check_entry(entry, block_size);
 955                 if (ret)
 956                         goto out;
 957                 data = buffer + entry->e_value_offs;
 958                 datalen = entry->e_value_size;
 959                 ret = copy_single_xattr(trans, root, objectid,
 960                                         entry, data, datalen);
 961                 if (ret)
 962                         goto out;
 963                 entry = EXT2_EXT_ATTR_NEXT(entry);
 964         }
 965 out:
 966         free(buffer);
 967         if ((void *)ext2_inode != inode_buf)
 968                 free(ext2_inode);
 969         return ret;
 970 }
 971 #define MINORBITS       20
 972 #define MKDEV(ma, mi)   (((ma) << MINORBITS) | (mi))
 973
 974 static inline dev_t old_decode_dev(u16 val)
 975 {
 976         return MKDEV((val >> 8) & 255, val & 255);
 977 }
 978
 979 static inline dev_t new_decode_dev(u32 dev)
 980 {
 981         unsigned major = (dev & 0xfff00) >> 8;
 982         unsigned minor = (dev & 0xff) | ((dev >> 12) & 0xfff00);
 983         return MKDEV(major, minor);
 984 }
 985
 986 static int copy_inode_item(struct btrfs_inode_item *dst,
 987                            struct ext2_inode *src, u32 blocksize)
 988 {
 989         btrfs_set_stack_inode_generation(dst, 1);
 990         btrfs_set_stack_inode_size(dst, src->i_size);
 991         btrfs_set_stack_inode_nbytes(dst, 0);
 992         btrfs_set_stack_inode_block_group(dst, 0);
 993         btrfs_set_stack_inode_nlink(dst, src->i_links_count);
 994         btrfs_set_stack_inode_uid(dst, src->i_uid | (src->i_uid_high << 16));
 995         btrfs_set_stack_inode_gid(dst, src->i_gid | (src->i_gid_high << 16));
 996         btrfs_set_stack_inode_mode(dst, src->i_mode);
 997         btrfs_set_stack_inode_rdev(dst, 0);
 998         btrfs_set_stack_inode_flags(dst, 0);
 999         btrfs_set_stack_timespec_sec(&dst->atime, src->i_atime);
1000         btrfs_set_stack_timespec_nsec(&dst->atime, 0);
1001         btrfs_set_stack_timespec_sec(&dst->ctime, src->i_ctime);
1002         btrfs_set_stack_timespec_nsec(&dst->ctime, 0);
1003         btrfs_set_stack_timespec_sec(&dst->mtime, src->i_mtime);
1004         btrfs_set_stack_timespec_nsec(&dst->mtime, 0);
1005         btrfs_set_stack_timespec_sec(&dst->otime, 0);
1006         btrfs_set_stack_timespec_nsec(&dst->otime, 0);
1007
1008         if (S_ISDIR(src->i_mode)) {
1009                 btrfs_set_stack_inode_size(dst, 0);
1010                 btrfs_set_stack_inode_nlink(dst, 1);
1011         }
1012         if (S_ISREG(src->i_mode)) {
1013                 btrfs_set_stack_inode_size(dst, (u64)src->i_size_high << 32 |
1014                                            (u64)src->i_size);
1015         }
1016         if (!S_ISREG(src->i_mode) && !S_ISDIR(src->i_mode) &&
1017             !S_ISLNK(src->i_mode)) {
1018                 if (src->i_block[0]) {
1019                         btrfs_set_stack_inode_rdev(dst,
1020                                 old_decode_dev(src->i_block[0]));
1021                 } else {
1022                         btrfs_set_stack_inode_rdev(dst,
1023                                 new_decode_dev(src->i_block[1]));
1024                 }
1025         }
1026         return 0;
1027 }
1028
1029 /*
1030  * copy a single inode. do all the required works, such as cloning
1031  * inode item, creating file extents and creating directory entries.
1032  */
1033 static int copy_single_inode(struct btrfs_trans_handle *trans,
1034                              struct btrfs_root *root, u64 objectid,
1035                              ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
1036                              struct ext2_inode *ext2_inode,
1037                              int datacsum, int packing, int noxattr)
1038 {
1039         int ret;
1040         struct btrfs_key inode_key;
1041         struct btrfs_inode_item btrfs_inode;
1042
1043         if (ext2_inode->i_links_count == 0)
1044                 return 0;
1045
1046         copy_inode_item(&btrfs_inode, ext2_inode, ext2_fs->blocksize);
1047         if (!datacsum && S_ISREG(ext2_inode->i_mode)) {
1048                 u32 flags = btrfs_stack_inode_flags(&btrfs_inode) |
1049                             BTRFS_INODE_NODATASUM;
1050                 btrfs_set_stack_inode_flags(&btrfs_inode, flags);
1051         }
1052
1053         switch (ext2_inode->i_mode & S_IFMT) {
1054         case S_IFREG:
1055                 ret = create_file_extents(trans, root, objectid, &btrfs_inode,
1056                                         ext2_fs, ext2_ino, datacsum, packing);
1057                 break;
1058         case S_IFDIR:
1059                 ret = create_dir_entries(trans, root, objectid, &btrfs_inode,
1060                                          ext2_fs, ext2_ino);
1061                 break;
1062         case S_IFLNK:
1063                 ret = create_symbol_link(trans, root, objectid, &btrfs_inode,
1064                                          ext2_fs, ext2_ino, ext2_inode);
1065                 break;
1066         default:
1067                 ret = 0;
1068                 break;
1069         }
1070         if (ret)
1071                 return ret;
1072
1073         if (!noxattr) {
1074                 ret = copy_extended_attrs(trans, root, objectid, &btrfs_inode,
1075                                           ext2_fs, ext2_ino);
1076                 if (ret)
1077                         return ret;
1078         }
1079         inode_key.objectid = objectid;
1080         inode_key.offset = 0;
1081         btrfs_set_key_type(&inode_key, BTRFS_INODE_ITEM_KEY);
1082         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1083         return ret;
1084 }
1085
1086 static int copy_disk_extent(struct btrfs_root *root, u64 dst_bytenr,
1087                             u64 src_bytenr, u32 num_bytes)
1088 {
1089         int ret;
1090         char *buffer;
1091         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
1092
1093         buffer = malloc(num_bytes);
1094         if (!buffer)
1095                 return -ENOMEM;
1096         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, src_bytenr);
1097         if (ret != num_bytes)
1098                 goto fail;
1099         ret = pwrite(fs_devs->latest_bdev, buffer, num_bytes, dst_bytenr);
1100         if (ret != num_bytes)
1101                 goto fail;
1102         ret = 0;
1103 fail:
1104         free(buffer);
1105         if (ret > 0)
1106                 ret = -1;
1107         return ret;
1108 }
1109 /*
1110  * scan ext2's inode bitmap and copy all used inodes.
1111  */
1112 static int copy_inodes(struct btrfs_root *root, ext2_filsys ext2_fs,
1113                        int datacsum, int packing, int noxattr, struct task_ctx *p)
1114 {
1115         int ret;
1116         errcode_t err;
1117         ext2_inode_scan ext2_scan;
1118         struct ext2_inode ext2_inode;
1119         ext2_ino_t ext2_ino;
1120         u64 objectid;
1121         struct btrfs_trans_handle *trans;
1122
1123         trans = btrfs_start_transaction(root, 1);
1124         if (!trans)
1125                 return -ENOMEM;
1126         err = ext2fs_open_inode_scan(ext2_fs, 0, &ext2_scan);
1127         if (err) {
1128                 fprintf(stderr, "ext2fs_open_inode_scan: %s\n", error_message(err));
1129                 return -1;
1130         }
1131         while (!(err = ext2fs_get_next_inode(ext2_scan, &ext2_ino,
1132                                              &ext2_inode))) {
1133                 /* no more inodes */
1134                 if (ext2_ino == 0)
1135                         break;
1136                 /* skip special inode in ext2fs */
1137                 if (ext2_ino < EXT2_GOOD_OLD_FIRST_INO &&
1138                     ext2_ino != EXT2_ROOT_INO)
1139                         continue;
1140                 objectid = ext2_ino + INO_OFFSET;
1141                 ret = copy_single_inode(trans, root,
1142                                         objectid, ext2_fs, ext2_ino,
1143                                         &ext2_inode, datacsum, packing,
1144                                         noxattr);
1145                 p->cur_copy_inodes++;
1146                 if (ret)
1147                         return ret;
1148                 if (trans->blocks_used >= 4096) {
1149                         ret = btrfs_commit_transaction(trans, root);
1150                         BUG_ON(ret);
1151                         trans = btrfs_start_transaction(root, 1);
1152                         BUG_ON(!trans);
1153                 }
1154         }
1155         if (err) {
1156                 fprintf(stderr, "ext2fs_get_next_inode: %s\n", error_message(err));
1157                 return -1;
1158         }
1159         ret = btrfs_commit_transaction(trans, root);
1160         BUG_ON(ret);
1161
1162         return ret;
1163 }
1164
1165 /*
1166  * Construct a range of ext2fs image file.
1167  * scan block allocation bitmap, find all blocks used by the ext2fs
1168  * in this range and create file extents that point to these blocks.
1169  *
1170  * Note: Before calling the function, no file extent points to blocks
1171  *       in this range
1172  */
1173 static int create_image_file_range(struct btrfs_trans_handle *trans,
1174                                    struct btrfs_root *root, u64 objectid,
1175                                    struct btrfs_inode_item *inode,
1176                                    u64 start_byte, u64 end_byte,
1177                                    ext2_filsys ext2_fs, int datacsum)
1178 {
1179         u32 blocksize = ext2_fs->blocksize;
1180         u32 block = start_byte / blocksize;
1181         u32 last_block = (end_byte + blocksize - 1) / blocksize;
1182         int ret = 0;
1183         struct blk_iterate_data data = {
1184                 .trans          = trans,
1185                 .root           = root,
1186                 .inode          = inode,
1187                 .objectid       = objectid,
1188                 .first_block    = block,
1189                 .disk_block     = 0,
1190                 .num_blocks     = 0,
1191                 .boundary       = (u64)-1,
1192                 .checksum       = datacsum,
1193                 .errcode        = 0,
1194         };
1195         for (; start_byte < end_byte; block++, start_byte += blocksize) {
1196                 if (!ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
1197                         continue;
1198                 ret = block_iterate_proc(NULL, block, block, &data);
1199                 if (ret & BLOCK_ABORT) {
1200                         ret = data.errcode;
1201                         goto fail;
1202                 }
1203         }
1204         if (data.num_blocks > 0) {
1205                 ret = record_file_blocks(trans, root, objectid, inode,
1206                                          data.first_block, data.disk_block,
1207                                          data.num_blocks, datacsum);
1208                 if (ret)
1209                         goto fail;
1210                 data.first_block += data.num_blocks;
1211         }
1212         if (last_block > data.first_block) {
1213                 ret = record_file_blocks(trans, root, objectid, inode,
1214                                          data.first_block, 0, last_block -
1215                                          data.first_block, datacsum);
1216                 if (ret)
1217                         goto fail;
1218         }
1219 fail:
1220         return ret;
1221 }
1222 /*
1223  * Create the ext2fs image file.
1224  */
1225 static int create_ext2_image(struct btrfs_root *root, ext2_filsys ext2_fs,
1226                              const char *name, int datacsum)
1227 {
1228         int ret;
1229         struct btrfs_key key;
1230         struct btrfs_key location;
1231         struct btrfs_path path;
1232         struct btrfs_inode_item btrfs_inode;
1233         struct btrfs_inode_item *inode_item;
1234         struct extent_buffer *leaf;
1235         struct btrfs_fs_info *fs_info = root->fs_info;
1236         struct btrfs_root *extent_root = fs_info->extent_root;
1237         struct btrfs_trans_handle *trans;
1238         struct btrfs_extent_item *ei;
1239         struct btrfs_extent_inline_ref *iref;
1240         struct btrfs_extent_data_ref *dref;
1241         u64 bytenr;
1242         u64 num_bytes;
1243         u64 objectid;
1244         u64 last_byte;
1245         u64 first_free;
1246         u64 total_bytes;
1247         u64 flags = BTRFS_INODE_READONLY;
1248         u32 sectorsize = root->sectorsize;
1249
1250         total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
1251         first_free =  BTRFS_SUPER_INFO_OFFSET + sectorsize * 2 - 1;
1252         first_free &= ~((u64)sectorsize - 1);
1253         if (!datacsum)
1254                 flags |= BTRFS_INODE_NODATASUM;
1255
1256         memset(&btrfs_inode, 0, sizeof(btrfs_inode));
1257         btrfs_set_stack_inode_generation(&btrfs_inode, 1);
1258         btrfs_set_stack_inode_size(&btrfs_inode, total_bytes);
1259         btrfs_set_stack_inode_nlink(&btrfs_inode, 1);
1260         btrfs_set_stack_inode_nbytes(&btrfs_inode, 0);
1261         btrfs_set_stack_inode_mode(&btrfs_inode, S_IFREG | 0400);
1262         btrfs_set_stack_inode_flags(&btrfs_inode,  flags);
1263         btrfs_init_path(&path);
1264         trans = btrfs_start_transaction(root, 1);
1265         BUG_ON(!trans);
1266
1267         objectid = btrfs_root_dirid(&root->root_item);
1268         ret = btrfs_find_free_objectid(trans, root, objectid, &objectid);
1269         if (ret)
1270                 goto fail;
1271
1272         /*
1273          * copy blocks covered by extent #0 to new positions. extent #0 is
1274          * special, we can't rely on relocate_extents_range to relocate it.
1275          */
1276         for (last_byte = 0; last_byte < first_free; last_byte += sectorsize) {
1277                 ret = custom_alloc_extent(root, sectorsize, 0, &key);
1278                 if (ret)
1279                         goto fail;
1280                 ret = copy_disk_extent(root, key.objectid, last_byte,
1281                                        sectorsize);
1282                 if (ret)
1283                         goto fail;
1284                 ret = btrfs_record_file_extent(trans, root, objectid,
1285                                                &btrfs_inode, last_byte,
1286                                                key.objectid, sectorsize);
1287                 if (ret)
1288                         goto fail;
1289                 if (datacsum) {
1290                         ret = csum_disk_extent(trans, root, key.objectid,
1291                                                sectorsize);
1292                         if (ret)
1293                                 goto fail;
1294                 }
1295         }
1296
1297         while(1) {
1298                 key.objectid = last_byte;
1299                 key.offset = 0;
1300                 btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
1301                 ret = btrfs_search_slot(trans, fs_info->extent_root,
1302                                         &key, &path, 0, 0);
1303                 if (ret < 0)
1304                         goto fail;
1305 next:
1306                 leaf = path.nodes[0];
1307                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1308                         ret = btrfs_next_leaf(extent_root, &path);
1309                         if (ret < 0)
1310                                 goto fail;
1311                         if (ret > 0)
1312                                 break;
1313                         leaf = path.nodes[0];
1314                 }
1315                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1316                 if (last_byte > key.objectid ||
1317                     key.type != BTRFS_EXTENT_ITEM_KEY) {
1318                         path.slots[0]++;
1319                         goto next;
1320                 }
1321
1322                 bytenr = key.objectid;
1323                 num_bytes = key.offset;
1324                 ei = btrfs_item_ptr(leaf, path.slots[0],
1325                                     struct btrfs_extent_item);
1326                 if (!(btrfs_extent_flags(leaf, ei) & BTRFS_EXTENT_FLAG_DATA)) {
1327                         path.slots[0]++;
1328                         goto next;
1329                 }
1330
1331                 BUG_ON(btrfs_item_size_nr(leaf, path.slots[0]) != sizeof(*ei) +
1332                        btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY));
1333
1334                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
1335                 key.type = btrfs_extent_inline_ref_type(leaf, iref);
1336                 BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
1337                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1338                 if (btrfs_extent_data_ref_root(leaf, dref) !=
1339                     BTRFS_FS_TREE_OBJECTID) {
1340                         path.slots[0]++;
1341                         goto next;
1342                 }
1343
1344                 if (bytenr > last_byte) {
1345                         ret = create_image_file_range(trans, root, objectid,
1346                                                       &btrfs_inode, last_byte,
1347                                                       bytenr, ext2_fs,
1348                                                       datacsum);
1349                         if (ret)
1350                                 goto fail;
1351                 }
1352                 ret = btrfs_record_file_extent(trans, root, objectid,
1353                                                &btrfs_inode, bytenr, bytenr,
1354                                                num_bytes);
1355                 if (ret)
1356                         goto fail;
1357                 last_byte = bytenr + num_bytes;
1358                 btrfs_release_path(&path);
1359
1360                 if (trans->blocks_used >= 4096) {
1361                         ret = btrfs_commit_transaction(trans, root);
1362                         BUG_ON(ret);
1363                         trans = btrfs_start_transaction(root, 1);
1364                         BUG_ON(!trans);
1365                 }
1366         }
1367         btrfs_release_path(&path);
1368         if (total_bytes > last_byte) {
1369                 ret = create_image_file_range(trans, root, objectid,
1370                                               &btrfs_inode, last_byte,
1371                                               total_bytes, ext2_fs,
1372                                               datacsum);
1373                 if (ret)
1374                         goto fail;
1375         }
1376
1377         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1378         if (ret)
1379                 goto fail;
1380
1381         location.objectid = objectid;
1382         location.offset = 0;
1383         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1384         ret = btrfs_insert_dir_item(trans, root, name, strlen(name),
1385                                     btrfs_root_dirid(&root->root_item),
1386                                     &location, EXT2_FT_REG_FILE, objectid);
1387         if (ret)
1388                 goto fail;
1389         ret = btrfs_insert_inode_ref(trans, root, name, strlen(name),
1390                                      objectid,
1391                                      btrfs_root_dirid(&root->root_item),
1392                                      objectid);
1393         if (ret)
1394                 goto fail;
1395         location.objectid = btrfs_root_dirid(&root->root_item);
1396         location.offset = 0;
1397         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1398         ret = btrfs_lookup_inode(trans, root, &path, &location, 1);
1399         if (ret)
1400                 goto fail;
1401         leaf = path.nodes[0];
1402         inode_item = btrfs_item_ptr(leaf, path.slots[0],
1403                                     struct btrfs_inode_item);
1404         btrfs_set_inode_size(leaf, inode_item, strlen(name) * 2 +
1405                              btrfs_inode_size(leaf, inode_item));
1406         btrfs_mark_buffer_dirty(leaf);
1407         btrfs_release_path(&path);
1408         ret = btrfs_commit_transaction(trans, root);
1409         BUG_ON(ret);
1410 fail:
1411         btrfs_release_path(&path);
1412         return ret;
1413 }
1414
1415 static struct btrfs_root * link_subvol(struct btrfs_root *root,
1416                 const char *base, u64 root_objectid)
1417 {
1418         struct btrfs_trans_handle *trans;
1419         struct btrfs_fs_info *fs_info = root->fs_info;
1420         struct btrfs_root *tree_root = fs_info->tree_root;
1421         struct btrfs_root *new_root = NULL;
1422         struct btrfs_path *path;
1423         struct btrfs_inode_item *inode_item;
1424         struct extent_buffer *leaf;
1425         struct btrfs_key key;
1426         u64 dirid = btrfs_root_dirid(&root->root_item);
1427         u64 index = 2;
1428         char buf[BTRFS_NAME_LEN + 1]; /* for snprintf null */
1429         int len;
1430         int i;
1431         int ret;
1432
1433         len = strlen(base);
1434         if (len < 1 || len > BTRFS_NAME_LEN)
1435                 return NULL;
1436
1437         path = btrfs_alloc_path();
1438         BUG_ON(!path);
1439
1440         key.objectid = dirid;
1441         key.type = BTRFS_DIR_INDEX_KEY;
1442         key.offset = (u64)-1;
1443
1444         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1445         BUG_ON(ret <= 0);
1446
1447         if (path->slots[0] > 0) {
1448                 path->slots[0]--;
1449                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1450                 if (key.objectid == dirid && key.type == BTRFS_DIR_INDEX_KEY)
1451                         index = key.offset + 1;
1452         }
1453         btrfs_release_path(path);
1454
1455         trans = btrfs_start_transaction(root, 1);
1456         BUG_ON(!trans);
1457
1458         key.objectid = dirid;
1459         key.offset = 0;
1460         key.type =  BTRFS_INODE_ITEM_KEY;
1461
1462         ret = btrfs_lookup_inode(trans, root, path, &key, 1);
1463         BUG_ON(ret);
1464         leaf = path->nodes[0];
1465         inode_item = btrfs_item_ptr(leaf, path->slots[0],
1466                                     struct btrfs_inode_item);
1467
1468         key.objectid = root_objectid;
1469         key.offset = (u64)-1;
1470         key.type = BTRFS_ROOT_ITEM_KEY;
1471
1472         memcpy(buf, base, len);
1473         for (i = 0; i < 1024; i++) {
1474                 ret = btrfs_insert_dir_item(trans, root, buf, len,
1475                                             dirid, &key, BTRFS_FT_DIR, index);
1476                 if (ret != -EEXIST)
1477                         break;
1478                 len = snprintf(buf, ARRAY_SIZE(buf), "%s%d", base, i);
1479                 if (len < 1 || len > BTRFS_NAME_LEN) {
1480                         ret = -EINVAL;
1481                         break;
1482                 }
1483         }
1484         if (ret)
1485                 goto fail;
1486
1487         btrfs_set_inode_size(leaf, inode_item, len * 2 +
1488                              btrfs_inode_size(leaf, inode_item));
1489         btrfs_mark_buffer_dirty(leaf);
1490         btrfs_release_path(path);
1491
1492         /* add the backref first */
1493         ret = btrfs_add_root_ref(trans, tree_root, root_objectid,
1494                                  BTRFS_ROOT_BACKREF_KEY,
1495                                  root->root_key.objectid,
1496                                  dirid, index, buf, len);
1497         BUG_ON(ret);
1498
1499         /* now add the forward ref */
1500         ret = btrfs_add_root_ref(trans, tree_root, root->root_key.objectid,
1501                                  BTRFS_ROOT_REF_KEY, root_objectid,
1502                                  dirid, index, buf, len);
1503
1504         ret = btrfs_commit_transaction(trans, root);
1505         BUG_ON(ret);
1506
1507         new_root = btrfs_read_fs_root(fs_info, &key);
1508         if (IS_ERR(new_root))
1509                 new_root = NULL;
1510 fail:
1511         btrfs_free_path(path);
1512         return new_root;
1513 }
1514
1515 static int create_chunk_mapping(struct btrfs_trans_handle *trans,
1516                                 struct btrfs_root *root)
1517 {
1518         struct btrfs_fs_info *info = root->fs_info;
1519         struct btrfs_root *chunk_root = info->chunk_root;
1520         struct btrfs_root *extent_root = info->extent_root;
1521         struct btrfs_device *device;
1522         struct btrfs_block_group_cache *cache;
1523         struct btrfs_dev_extent *extent;
1524         struct extent_buffer *leaf;
1525         struct btrfs_chunk chunk;
1526         struct btrfs_key key;
1527         struct btrfs_path path;
1528         u64 cur_start;
1529         u64 total_bytes;
1530         u64 chunk_objectid;
1531         int ret;
1532
1533         btrfs_init_path(&path);
1534
1535         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
1536         chunk_objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
1537
1538         BUG_ON(list_empty(&info->fs_devices->devices));
1539         device = list_entry(info->fs_devices->devices.next,
1540                             struct btrfs_device, dev_list);
1541         BUG_ON(device->devid != info->fs_devices->latest_devid);
1542
1543         /* delete device extent created by make_btrfs */
1544         key.objectid = device->devid;
1545         key.offset = 0;
1546         key.type = BTRFS_DEV_EXTENT_KEY;
1547         ret = btrfs_search_slot(trans, device->dev_root, &key, &path, -1, 1);
1548         if (ret < 0)
1549                 goto err;
1550
1551         BUG_ON(ret > 0);
1552         ret = btrfs_del_item(trans, device->dev_root, &path);
1553         if (ret)
1554                 goto err;
1555         btrfs_release_path(&path);
1556
1557         /* delete chunk item created by make_btrfs */
1558         key.objectid = chunk_objectid;
1559         key.offset = 0;
1560         key.type = BTRFS_CHUNK_ITEM_KEY;
1561         ret = btrfs_search_slot(trans, chunk_root, &key, &path, -1, 1);
1562         if (ret < 0)
1563                 goto err;
1564
1565         BUG_ON(ret > 0);
1566         ret = btrfs_del_item(trans, chunk_root, &path);
1567         if (ret)
1568                 goto err;
1569         btrfs_release_path(&path);
1570
1571         /* for each block group, create device extent and chunk item */
1572         cur_start = 0;
1573         while (cur_start < total_bytes) {
1574                 cache = btrfs_lookup_block_group(root->fs_info, cur_start);
1575                 BUG_ON(!cache);
1576
1577                 /* insert device extent */
1578                 key.objectid = device->devid;
1579                 key.offset = cache->key.objectid;
1580                 key.type = BTRFS_DEV_EXTENT_KEY;
1581                 ret = btrfs_insert_empty_item(trans, device->dev_root, &path,
1582                                               &key, sizeof(*extent));
1583                 if (ret)
1584                         goto err;
1585
1586                 leaf = path.nodes[0];
1587                 extent = btrfs_item_ptr(leaf, path.slots[0],
1588                                         struct btrfs_dev_extent);
1589
1590                 btrfs_set_dev_extent_chunk_tree(leaf, extent,
1591                                                 chunk_root->root_key.objectid);
1592                 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
1593                                                     chunk_objectid);
1594                 btrfs_set_dev_extent_chunk_offset(leaf, extent,
1595                                                   cache->key.objectid);
1596                 btrfs_set_dev_extent_length(leaf, extent, cache->key.offset);
1597                 write_extent_buffer(leaf, root->fs_info->chunk_tree_uuid,
1598                     (unsigned long)btrfs_dev_extent_chunk_tree_uuid(extent),
1599                     BTRFS_UUID_SIZE);
1600                 btrfs_mark_buffer_dirty(leaf);
1601                 btrfs_release_path(&path);
1602
1603                 /* insert chunk item */
1604                 btrfs_set_stack_chunk_length(&chunk, cache->key.offset);
1605                 btrfs_set_stack_chunk_owner(&chunk,
1606                                             extent_root->root_key.objectid);
1607                 btrfs_set_stack_chunk_stripe_len(&chunk, BTRFS_STRIPE_LEN);
1608                 btrfs_set_stack_chunk_type(&chunk, cache->flags);
1609                 btrfs_set_stack_chunk_io_align(&chunk, device->io_align);
1610                 btrfs_set_stack_chunk_io_width(&chunk, device->io_width);
1611                 btrfs_set_stack_chunk_sector_size(&chunk, device->sector_size);
1612                 btrfs_set_stack_chunk_num_stripes(&chunk, 1);
1613                 btrfs_set_stack_chunk_sub_stripes(&chunk, 0);
1614                 btrfs_set_stack_stripe_devid(&chunk.stripe, device->devid);
1615                 btrfs_set_stack_stripe_offset(&chunk.stripe,
1616                                               cache->key.objectid);
1617                 memcpy(&chunk.stripe.dev_uuid, device->uuid, BTRFS_UUID_SIZE);
1618
1619                 key.objectid = chunk_objectid;
1620                 key.offset = cache->key.objectid;
1621                 key.type = BTRFS_CHUNK_ITEM_KEY;
1622
1623                 ret = btrfs_insert_item(trans, chunk_root, &key, &chunk,
1624                                         btrfs_chunk_item_size(1));
1625                 if (ret)
1626                         goto err;
1627
1628                 cur_start = cache->key.objectid + cache->key.offset;
1629         }
1630
1631         device->bytes_used = total_bytes;
1632         ret = btrfs_update_device(trans, device);
1633 err:
1634         btrfs_release_path(&path);
1635         return ret;
1636 }
1637
1638 static int create_subvol(struct btrfs_trans_handle *trans,
1639                          struct btrfs_root *root, u64 root_objectid)
1640 {
1641         struct extent_buffer *tmp;
1642         struct btrfs_root *new_root;
1643         struct btrfs_key key;
1644         struct btrfs_root_item root_item;
1645         int ret;
1646
1647         ret = btrfs_copy_root(trans, root, root->node, &tmp,
1648                               root_objectid);
1649         BUG_ON(ret);
1650
1651         memcpy(&root_item, &root->root_item, sizeof(root_item));
1652         btrfs_set_root_bytenr(&root_item, tmp->start);
1653         btrfs_set_root_level(&root_item, btrfs_header_level(tmp));
1654         btrfs_set_root_generation(&root_item, trans->transid);
1655         free_extent_buffer(tmp);
1656
1657         key.objectid = root_objectid;
1658         key.type = BTRFS_ROOT_ITEM_KEY;
1659         key.offset = trans->transid;
1660         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1661                                 &key, &root_item);
1662
1663         key.offset = (u64)-1;
1664         new_root = btrfs_read_fs_root(root->fs_info, &key);
1665         BUG_ON(!new_root || IS_ERR(new_root));
1666
1667         ret = btrfs_make_root_dir(trans, new_root, BTRFS_FIRST_FREE_OBJECTID);
1668         BUG_ON(ret);
1669
1670         return 0;
1671 }
1672
1673 static int init_btrfs(struct btrfs_root *root)
1674 {
1675         int ret;
1676         struct btrfs_key location;
1677         struct btrfs_trans_handle *trans;
1678         struct btrfs_fs_info *fs_info = root->fs_info;
1679         struct extent_buffer *tmp;
1680
1681         trans = btrfs_start_transaction(root, 1);
1682         BUG_ON(!trans);
1683         ret = btrfs_make_block_groups(trans, root);
1684         if (ret)
1685                 goto err;
1686         ret = btrfs_fix_block_accounting(trans, root);
1687         if (ret)
1688                 goto err;
1689         ret = create_chunk_mapping(trans, root);
1690         if (ret)
1691                 goto err;
1692         ret = btrfs_make_root_dir(trans, fs_info->tree_root,
1693                                   BTRFS_ROOT_TREE_DIR_OBJECTID);
1694         if (ret)
1695                 goto err;
1696         memcpy(&location, &root->root_key, sizeof(location));
1697         location.offset = (u64)-1;
1698         ret = btrfs_insert_dir_item(trans, fs_info->tree_root, "default", 7,
1699                                 btrfs_super_root_dir(fs_info->super_copy),
1700                                 &location, BTRFS_FT_DIR, 0);
1701         if (ret)
1702                 goto err;
1703         ret = btrfs_insert_inode_ref(trans, fs_info->tree_root, "default", 7,
1704                                 location.objectid,
1705                                 btrfs_super_root_dir(fs_info->super_copy), 0);
1706         if (ret)
1707                 goto err;
1708         btrfs_set_root_dirid(&fs_info->fs_root->root_item,
1709                              BTRFS_FIRST_FREE_OBJECTID);
1710
1711         /* subvol for ext2 image file */
1712         ret = create_subvol(trans, root, EXT2_IMAGE_SUBVOL_OBJECTID);
1713         BUG_ON(ret);
1714         /* subvol for data relocation */
1715         ret = create_subvol(trans, root, BTRFS_DATA_RELOC_TREE_OBJECTID);
1716         BUG_ON(ret);
1717
1718         extent_buffer_get(fs_info->csum_root->node);
1719         ret = __btrfs_cow_block(trans, fs_info->csum_root,
1720                                 fs_info->csum_root->node, NULL, 0, &tmp, 0, 0);
1721         BUG_ON(ret);
1722         free_extent_buffer(tmp);
1723
1724         ret = btrfs_commit_transaction(trans, root);
1725         BUG_ON(ret);
1726 err:
1727         return ret;
1728 }
1729
1730 /*
1731  * Migrate super block to its default position and zero 0 ~ 16k
1732  */
1733 static int migrate_super_block(int fd, u64 old_bytenr, u32 sectorsize)
1734 {
1735         int ret;
1736         struct extent_buffer *buf;
1737         struct btrfs_super_block *super;
1738         u32 len;
1739         u32 bytenr;
1740
1741         BUG_ON(sectorsize < sizeof(*super));
1742         buf = malloc(sizeof(*buf) + sectorsize);
1743         if (!buf)
1744                 return -ENOMEM;
1745
1746         buf->len = sectorsize;
1747         ret = pread(fd, buf->data, sectorsize, old_bytenr);
1748         if (ret != sectorsize)
1749                 goto fail;
1750
1751         super = (struct btrfs_super_block *)buf->data;
1752         BUG_ON(btrfs_super_bytenr(super) != old_bytenr);
1753         btrfs_set_super_bytenr(super, BTRFS_SUPER_INFO_OFFSET);
1754
1755         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1756         ret = pwrite(fd, buf->data, sectorsize, BTRFS_SUPER_INFO_OFFSET);
1757         if (ret != sectorsize)
1758                 goto fail;
1759
1760         ret = fsync(fd);
1761         if (ret)
1762                 goto fail;
1763
1764         memset(buf->data, 0, sectorsize);
1765         for (bytenr = 0; bytenr < BTRFS_SUPER_INFO_OFFSET; ) {
1766                 len = BTRFS_SUPER_INFO_OFFSET - bytenr;
1767                 if (len > sectorsize)
1768                         len = sectorsize;
1769                 ret = pwrite(fd, buf->data, len, bytenr);
1770                 if (ret != len) {
1771                         fprintf(stderr, "unable to zero fill device\n");
1772                         break;
1773                 }
1774                 bytenr += len;
1775         }
1776         ret = 0;
1777         fsync(fd);
1778 fail:
1779         free(buf);
1780         if (ret > 0)
1781                 ret = -1;
1782         return ret;
1783 }
1784
1785 static int prepare_system_chunk_sb(struct btrfs_super_block *super)
1786 {
1787         struct btrfs_chunk *chunk;
1788         struct btrfs_disk_key *key;
1789         u32 sectorsize = btrfs_super_sectorsize(super);
1790
1791         key = (struct btrfs_disk_key *)(super->sys_chunk_array);
1792         chunk = (struct btrfs_chunk *)(super->sys_chunk_array +
1793                                        sizeof(struct btrfs_disk_key));
1794
1795         btrfs_set_disk_key_objectid(key, BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1796         btrfs_set_disk_key_type(key, BTRFS_CHUNK_ITEM_KEY);
1797         btrfs_set_disk_key_offset(key, 0);
1798
1799         btrfs_set_stack_chunk_length(chunk, btrfs_super_total_bytes(super));
1800         btrfs_set_stack_chunk_owner(chunk, BTRFS_EXTENT_TREE_OBJECTID);
1801         btrfs_set_stack_chunk_stripe_len(chunk, BTRFS_STRIPE_LEN);
1802         btrfs_set_stack_chunk_type(chunk, BTRFS_BLOCK_GROUP_SYSTEM);
1803         btrfs_set_stack_chunk_io_align(chunk, sectorsize);
1804         btrfs_set_stack_chunk_io_width(chunk, sectorsize);
1805         btrfs_set_stack_chunk_sector_size(chunk, sectorsize);
1806         btrfs_set_stack_chunk_num_stripes(chunk, 1);
1807         btrfs_set_stack_chunk_sub_stripes(chunk, 0);
1808         chunk->stripe.devid = super->dev_item.devid;
1809         btrfs_set_stack_stripe_offset(&chunk->stripe, 0);
1810         memcpy(chunk->stripe.dev_uuid, super->dev_item.uuid, BTRFS_UUID_SIZE);
1811         btrfs_set_super_sys_array_size(super, sizeof(*key) + sizeof(*chunk));
1812         return 0;
1813 }
1814
1815 static int prepare_system_chunk(int fd, u64 sb_bytenr)
1816 {
1817         int ret;
1818         struct extent_buffer *buf;
1819         struct btrfs_super_block *super;
1820
1821         BUG_ON(BTRFS_SUPER_INFO_SIZE < sizeof(*super));
1822         buf = malloc(sizeof(*buf) + BTRFS_SUPER_INFO_SIZE);
1823         if (!buf)
1824                 return -ENOMEM;
1825
1826         buf->len = BTRFS_SUPER_INFO_SIZE;
1827         ret = pread(fd, buf->data, BTRFS_SUPER_INFO_SIZE, sb_bytenr);
1828         if (ret != BTRFS_SUPER_INFO_SIZE)
1829                 goto fail;
1830
1831         super = (struct btrfs_super_block *)buf->data;
1832         BUG_ON(btrfs_super_bytenr(super) != sb_bytenr);
1833         BUG_ON(btrfs_super_num_devices(super) != 1);
1834
1835         ret = prepare_system_chunk_sb(super);
1836         if (ret)
1837                 goto fail;
1838
1839         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1840         ret = pwrite(fd, buf->data, BTRFS_SUPER_INFO_SIZE, sb_bytenr);
1841         if (ret != BTRFS_SUPER_INFO_SIZE)
1842                 goto fail;
1843
1844         ret = 0;
1845 fail:
1846         free(buf);
1847         if (ret > 0)
1848                 ret = -1;
1849         return ret;
1850 }
1851
1852 static int relocate_one_reference(struct btrfs_trans_handle *trans,
1853                                   struct btrfs_root *root,
1854                                   u64 extent_start, u64 extent_size,
1855                                   struct btrfs_key *extent_key,
1856                                   struct extent_io_tree *reloc_tree)
1857 {
1858         struct extent_buffer *leaf;
1859         struct btrfs_file_extent_item *fi;
1860         struct btrfs_key key;
1861         struct btrfs_path path;
1862         struct btrfs_inode_item inode;
1863         struct blk_iterate_data data;
1864         u64 bytenr;
1865         u64 num_bytes;
1866         u64 cur_offset;
1867         u64 new_pos;
1868         u64 nbytes;
1869         u64 sector_end;
1870         u32 sectorsize = root->sectorsize;
1871         unsigned long ptr;
1872         int datacsum;
1873         int fd;
1874         int ret;
1875
1876         btrfs_init_path(&path);
1877         ret = btrfs_search_slot(trans, root, extent_key, &path, -1, 1);
1878         if (ret)
1879                 goto fail;
1880
1881         leaf = path.nodes[0];
1882         fi = btrfs_item_ptr(leaf, path.slots[0],
1883                             struct btrfs_file_extent_item);
1884         BUG_ON(btrfs_file_extent_offset(leaf, fi) > 0);
1885         if (extent_start != btrfs_file_extent_disk_bytenr(leaf, fi) ||
1886             extent_size != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1887                 ret = 1;
1888                 goto fail;
1889         }
1890
1891         bytenr = extent_start + btrfs_file_extent_offset(leaf, fi);
1892         num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
1893
1894         ret = btrfs_del_item(trans, root, &path);
1895         if (ret)
1896                 goto fail;
1897
1898         ret = btrfs_free_extent(trans, root, extent_start, extent_size, 0,
1899                                 root->root_key.objectid,
1900                                 extent_key->objectid, extent_key->offset);
1901         if (ret)
1902                 goto fail;
1903
1904         btrfs_release_path(&path);
1905
1906         key.objectid = extent_key->objectid;
1907         key.offset = 0;
1908         key.type =  BTRFS_INODE_ITEM_KEY;
1909         ret = btrfs_lookup_inode(trans, root, &path, &key, 0);
1910         if (ret)
1911                 goto fail;
1912
1913         leaf = path.nodes[0];
1914         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1915         read_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1916         btrfs_release_path(&path);
1917
1918         BUG_ON(num_bytes & (sectorsize - 1));
1919         nbytes = btrfs_stack_inode_nbytes(&inode) - num_bytes;
1920         btrfs_set_stack_inode_nbytes(&inode, nbytes);
1921         datacsum = !(btrfs_stack_inode_flags(&inode) & BTRFS_INODE_NODATASUM);
1922
1923         data = (struct blk_iterate_data) {
1924                 .trans          = trans,
1925                 .root           = root,
1926                 .inode          = &inode,
1927                 .objectid       = extent_key->objectid,
1928                 .first_block    = extent_key->offset / sectorsize,
1929                 .disk_block     = 0,
1930                 .num_blocks     = 0,
1931                 .boundary       = (u64)-1,
1932                 .checksum       = datacsum,
1933                 .errcode        = 0,
1934         };
1935
1936         cur_offset = extent_key->offset;
1937         while (num_bytes > 0) {
1938                 sector_end = bytenr + sectorsize - 1;
1939                 if (test_range_bit(reloc_tree, bytenr, sector_end,
1940                                    EXTENT_LOCKED, 1)) {
1941                         ret = get_state_private(reloc_tree, bytenr, &new_pos);
1942                         BUG_ON(ret);
1943                 } else {
1944                         ret = custom_alloc_extent(root, sectorsize, 0, &key);
1945                         if (ret)
1946                                 goto fail;
1947                         new_pos = key.objectid;
1948
1949                         if (cur_offset == extent_key->offset) {
1950                                 fd = root->fs_info->fs_devices->latest_bdev;
1951                                 readahead(fd, bytenr, num_bytes);
1952                         }
1953                         ret = copy_disk_extent(root, new_pos, bytenr,
1954                                                sectorsize);
1955                         if (ret)
1956                                 goto fail;
1957                         ret = set_extent_bits(reloc_tree, bytenr, sector_end,
1958                                               EXTENT_LOCKED, GFP_NOFS);
1959                         BUG_ON(ret);
1960                         ret = set_state_private(reloc_tree, bytenr, new_pos);
1961                         BUG_ON(ret);
1962                 }
1963
1964                 ret = block_iterate_proc(NULL, new_pos / sectorsize,
1965                                          cur_offset / sectorsize, &data);
1966                 if (ret & BLOCK_ABORT) {
1967                         ret = data.errcode;
1968                         goto fail;
1969                 }
1970
1971                 cur_offset += sectorsize;
1972                 bytenr += sectorsize;
1973                 num_bytes -= sectorsize;
1974         }
1975
1976         if (data.num_blocks > 0) {
1977                 ret = record_file_blocks(trans, root,
1978                                          extent_key->objectid, &inode,
1979                                          data.first_block, data.disk_block,
1980                                          data.num_blocks, datacsum);
1981                 if (ret)
1982                         goto fail;
1983         }
1984
1985         key.objectid = extent_key->objectid;
1986         key.offset = 0;
1987         key.type =  BTRFS_INODE_ITEM_KEY;
1988         ret = btrfs_lookup_inode(trans, root, &path, &key, 1);
1989         if (ret)
1990                 goto fail;
1991
1992         leaf = path.nodes[0];
1993         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1994         write_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1995         btrfs_mark_buffer_dirty(leaf);
1996         btrfs_release_path(&path);
1997
1998 fail:
1999         btrfs_release_path(&path);
2000         return ret;
2001 }
2002
2003 static int relocate_extents_range(struct btrfs_root *fs_root,
2004                                   struct btrfs_root *ext2_root,
2005                                   u64 start_byte, u64 end_byte)
2006 {
2007         struct btrfs_fs_info *info = fs_root->fs_info;
2008         struct btrfs_root *extent_root = info->extent_root;
2009         struct btrfs_root *cur_root = NULL;
2010         struct btrfs_trans_handle *trans;
2011         struct btrfs_extent_data_ref *dref;
2012         struct btrfs_extent_inline_ref *iref;
2013         struct btrfs_extent_item *ei;
2014         struct extent_buffer *leaf;
2015         struct btrfs_key key;
2016         struct btrfs_key extent_key;
2017         struct btrfs_path path;
2018         struct extent_io_tree reloc_tree;
2019         unsigned long ptr;
2020         unsigned long end;
2021         u64 cur_byte;
2022         u64 num_bytes;
2023         u64 ref_root;
2024         u64 num_extents;
2025         int pass = 0;
2026         int ret;
2027
2028         btrfs_init_path(&path);
2029         extent_io_tree_init(&reloc_tree);
2030
2031         key.objectid = start_byte;
2032         key.offset = 0;
2033         key.type = BTRFS_EXTENT_ITEM_KEY;
2034         ret = btrfs_search_slot(NULL, extent_root, &key, &path, 0, 0);
2035         if (ret < 0)
2036                 goto fail;
2037         if (ret > 0) {
2038                 ret = btrfs_previous_item(extent_root, &path, 0,
2039                                           BTRFS_EXTENT_ITEM_KEY);
2040                 if (ret < 0)
2041                         goto fail;
2042                 if (ret == 0) {
2043                         leaf = path.nodes[0];
2044                         btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2045                         if (key.objectid + key.offset > start_byte)
2046                                 start_byte = key.objectid;
2047                 }
2048         }
2049         btrfs_release_path(&path);
2050 again:
2051         cur_root = (pass % 2 == 0) ? ext2_root : fs_root;
2052         num_extents = 0;
2053
2054         trans = btrfs_start_transaction(cur_root, 1);
2055         BUG_ON(!trans);
2056
2057         cur_byte = start_byte;
2058         while (1) {
2059                 key.objectid = cur_byte;
2060                 key.offset = 0;
2061                 key.type = BTRFS_EXTENT_ITEM_KEY;
2062                 ret = btrfs_search_slot(trans, extent_root,
2063                                         &key, &path, 0, 0);
2064                 if (ret < 0)
2065                         goto fail;
2066 next:
2067                 leaf = path.nodes[0];
2068                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2069                         ret = btrfs_next_leaf(extent_root, &path);
2070                         if (ret < 0)
2071                                 goto fail;
2072                         if (ret > 0)
2073                                 break;
2074                         leaf = path.nodes[0];
2075                 }
2076
2077                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2078                 if (key.objectid < cur_byte ||
2079                     key.type != BTRFS_EXTENT_ITEM_KEY) {
2080                         path.slots[0]++;
2081                         goto next;
2082                 }
2083                 if (key.objectid >= end_byte)
2084                         break;
2085
2086                 num_extents++;
2087
2088                 cur_byte = key.objectid;
2089                 num_bytes = key.offset;
2090                 ei = btrfs_item_ptr(leaf, path.slots[0],
2091                                     struct btrfs_extent_item);
2092                 BUG_ON(!(btrfs_extent_flags(leaf, ei) &
2093                          BTRFS_EXTENT_FLAG_DATA));
2094
2095                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2096                 end = ptr + btrfs_item_size_nr(leaf, path.slots[0]);
2097
2098                 ptr += sizeof(struct btrfs_extent_item);
2099
2100                 while (ptr < end) {
2101                         iref = (struct btrfs_extent_inline_ref *)ptr;
2102                         key.type = btrfs_extent_inline_ref_type(leaf, iref);
2103                         BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
2104                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2105                         ref_root = btrfs_extent_data_ref_root(leaf, dref);
2106                         extent_key.objectid =
2107                                 btrfs_extent_data_ref_objectid(leaf, dref);
2108                         extent_key.offset =
2109                                 btrfs_extent_data_ref_offset(leaf, dref);
2110                         extent_key.type = BTRFS_EXTENT_DATA_KEY;
2111                         BUG_ON(btrfs_extent_data_ref_count(leaf, dref) != 1);
2112
2113                         if (ref_root == cur_root->root_key.objectid)
2114                                 break;
2115
2116                         ptr += btrfs_extent_inline_ref_size(key.type);
2117                 }
2118
2119                 if (ptr >= end) {
2120                         path.slots[0]++;
2121                         goto next;
2122                 }
2123
2124                 ret = relocate_one_reference(trans, cur_root, cur_byte,
2125                                              num_bytes, &extent_key,
2126                                              &reloc_tree);
2127                 if (ret < 0)
2128                         goto fail;
2129
2130                 cur_byte += num_bytes;
2131                 btrfs_release_path(&path);
2132
2133                 if (trans->blocks_used >= 4096) {
2134                         ret = btrfs_commit_transaction(trans, cur_root);
2135                         BUG_ON(ret);
2136                         trans = btrfs_start_transaction(cur_root, 1);
2137                         BUG_ON(!trans);
2138                 }
2139         }
2140         btrfs_release_path(&path);
2141
2142         ret = btrfs_commit_transaction(trans, cur_root);
2143         BUG_ON(ret);
2144
2145         if (num_extents > 0 && pass++ < 16)
2146                 goto again;
2147
2148         ret = (num_extents > 0) ? -1 : 0;
2149 fail:
2150         btrfs_release_path(&path);
2151         extent_io_tree_cleanup(&reloc_tree);
2152         return ret;
2153 }
2154
2155 /*
2156  * relocate data in system chunk
2157  */
2158 static int cleanup_sys_chunk(struct btrfs_root *fs_root,
2159                              struct btrfs_root *ext2_root)
2160 {
2161         struct btrfs_block_group_cache *cache;
2162         int i, ret = 0;
2163         u64 offset = 0;
2164         u64 end_byte;
2165
2166         while(1) {
2167                 cache = btrfs_lookup_block_group(fs_root->fs_info, offset);
2168                 if (!cache)
2169                         break;
2170
2171                 end_byte = cache->key.objectid + cache->key.offset;
2172                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2173                         ret = relocate_extents_range(fs_root, ext2_root,
2174                                                      cache->key.objectid,
2175                                                      end_byte);
2176                         if (ret)
2177                                 goto fail;
2178                 }
2179                 offset = end_byte;
2180         }
2181         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2182                 offset = btrfs_sb_offset(i);
2183                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
2184
2185                 ret = relocate_extents_range(fs_root, ext2_root,
2186                                              offset, offset + BTRFS_STRIPE_LEN);
2187                 if (ret)
2188                         goto fail;
2189         }
2190         ret = 0;
2191 fail:
2192         return ret;
2193 }
2194
2195 static int fixup_chunk_mapping(struct btrfs_root *root)
2196 {
2197         struct btrfs_trans_handle *trans;
2198         struct btrfs_fs_info *info = root->fs_info;
2199         struct btrfs_root *chunk_root = info->chunk_root;
2200         struct extent_buffer *leaf;
2201         struct btrfs_key key;
2202         struct btrfs_path path;
2203         struct btrfs_chunk chunk;
2204         unsigned long ptr;
2205         u32 size;
2206         u64 type;
2207         int ret;
2208
2209         btrfs_init_path(&path);
2210
2211         trans = btrfs_start_transaction(root, 1);
2212         BUG_ON(!trans);
2213
2214         /*
2215          * recow the whole chunk tree. this will move all chunk tree blocks
2216          * into system block group.
2217          */
2218         memset(&key, 0, sizeof(key));
2219         while (1) {
2220                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2221                 if (ret < 0)
2222                         goto err;
2223
2224                 ret = btrfs_next_leaf(chunk_root, &path);
2225                 if (ret < 0)
2226                         goto err;
2227                 if (ret > 0)
2228                         break;
2229
2230                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2231                 btrfs_release_path(&path);
2232         }
2233         btrfs_release_path(&path);
2234
2235         /* fixup the system chunk array in super block */
2236         btrfs_set_super_sys_array_size(info->super_copy, 0);
2237
2238         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2239         key.offset = 0;
2240         key.type = BTRFS_CHUNK_ITEM_KEY;
2241
2242         ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 0);
2243         if (ret < 0)
2244                 goto err;
2245         BUG_ON(ret != 0);
2246         while(1) {
2247                 leaf = path.nodes[0];
2248                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2249                         ret = btrfs_next_leaf(chunk_root, &path);
2250                         if (ret < 0)
2251                                 goto err;
2252                         if (ret > 0)
2253                                 break;
2254                         leaf = path.nodes[0];
2255                 }
2256                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2257                 if (key.type != BTRFS_CHUNK_ITEM_KEY)
2258                         goto next;
2259
2260                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2261                 size = btrfs_item_size_nr(leaf, path.slots[0]);
2262                 BUG_ON(size != sizeof(chunk));
2263                 read_extent_buffer(leaf, &chunk, ptr, size);
2264                 type = btrfs_stack_chunk_type(&chunk);
2265
2266                 if (!(type & BTRFS_BLOCK_GROUP_SYSTEM))
2267                         goto next;
2268
2269                 ret = btrfs_add_system_chunk(trans, chunk_root, &key,
2270                                              &chunk, size);
2271                 if (ret)
2272                         goto err;
2273 next:
2274                 path.slots[0]++;
2275         }
2276
2277         ret = btrfs_commit_transaction(trans, root);
2278         BUG_ON(ret);
2279 err:
2280         btrfs_release_path(&path);
2281         return ret;
2282 }
2283
2284 static int do_convert(const char *devname, int datacsum, int packing, int noxattr,
2285                 u32 nodesize, int copylabel, const char *fslabel, int progress)
2286 {
2287         int i, ret, blocks_per_node;
2288         int fd = -1;
2289         u32 blocksize;
2290         u64 blocks[7];
2291         u64 total_bytes;
2292         u64 super_bytenr;
2293         ext2_filsys ext2_fs;
2294         struct btrfs_root *root;
2295         struct btrfs_root *ext2_root;
2296         struct task_ctx ctx;
2297
2298         ret = open_ext2fs(devname, &ext2_fs);
2299         if (ret) {
2300                 fprintf(stderr, "unable to open the Ext2fs\n");
2301                 goto fail;
2302         }
2303         blocksize = ext2_fs->blocksize;
2304         total_bytes = (u64)ext2_fs->super->s_blocks_count * blocksize;
2305         if (blocksize < 4096) {
2306                 fprintf(stderr, "block size is too small\n");
2307                 goto fail;
2308         }
2309         if (!(ext2_fs->super->s_feature_incompat &
2310               EXT2_FEATURE_INCOMPAT_FILETYPE)) {
2311                 fprintf(stderr, "filetype feature is missing\n");
2312                 goto fail;
2313         }
2314         if (btrfs_check_nodesize(nodesize, blocksize))
2315                 goto fail;
2316         blocks_per_node = nodesize / blocksize;
2317         ret = -blocks_per_node;
2318         for (i = 0; i < 7; i++) {
2319                 if (nodesize == blocksize)
2320                         ret = ext2_alloc_block(ext2_fs, 0, blocks + i);
2321                 else
2322                         ret = ext2_alloc_block_range(ext2_fs,
2323                                         ret + blocks_per_node, blocks_per_node,
2324                                         blocks + i);
2325                 if (ret) {
2326                         fprintf(stderr, "not enough free space\n");
2327                         goto fail;
2328                 }
2329                 blocks[i] *= blocksize;
2330         }
2331         super_bytenr = blocks[0];
2332         fd = open(devname, O_RDWR);
2333         if (fd < 0) {
2334                 fprintf(stderr, "unable to open %s\n", devname);
2335                 goto fail;
2336         }
2337         ret = make_btrfs(fd, devname, ext2_fs->super->s_volume_name,
2338                          NULL, blocks, total_bytes, nodesize,
2339                          blocksize, blocksize, 0);
2340         if (ret) {
2341                 fprintf(stderr, "unable to create initial ctree: %s\n",
2342                         strerror(-ret));
2343                 goto fail;
2344         }
2345         /* create a system chunk that maps the whole device */
2346         ret = prepare_system_chunk(fd, super_bytenr);
2347         if (ret) {
2348                 fprintf(stderr, "unable to update system chunk\n");
2349                 goto fail;
2350         }
2351         root = open_ctree_fd(fd, devname, super_bytenr, OPEN_CTREE_WRITES);
2352         if (!root) {
2353                 fprintf(stderr, "unable to open ctree\n");
2354                 goto fail;
2355         }
2356         ret = cache_free_extents(root, ext2_fs);
2357         if (ret) {
2358                 fprintf(stderr, "error during cache_free_extents %d\n", ret);
2359                 goto fail;
2360         }
2361         root->fs_info->extent_ops = &extent_ops;
2362         /* recover block allocation bitmap */
2363         for (i = 0; i < 7; i++) {
2364                 blocks[i] /= blocksize;
2365                 if (nodesize == blocksize)
2366                         ext2_free_block(ext2_fs, blocks[i]);
2367                 else
2368                         ext2_free_block_range(ext2_fs, blocks[i],
2369                                         blocks_per_node);
2370         }
2371         ret = init_btrfs(root);
2372         if (ret) {
2373                 fprintf(stderr, "unable to setup the root tree\n");
2374                 goto fail;
2375         }
2376         printf("creating btrfs metadata.\n");
2377         ctx.max_copy_inodes = (ext2_fs->super->s_inodes_count
2378                         - ext2_fs->super->s_free_inodes_count);
2379         ctx.cur_copy_inodes = 0;
2380
2381         if (progress) {
2382                 ctx.info = task_init(print_copied_inodes, after_copied_inodes, &ctx);
2383                 task_start(ctx.info);
2384         }
2385         ret = copy_inodes(root, ext2_fs, datacsum, packing, noxattr, &ctx);
2386         if (ret) {
2387                 fprintf(stderr, "error during copy_inodes %d\n", ret);
2388                 goto fail;
2389         }
2390         if (progress) {
2391                 task_stop(ctx.info);
2392                 task_deinit(ctx.info);
2393         }
2394         printf("creating ext2fs image file.\n");
2395         ext2_root = link_subvol(root, "ext2_saved", EXT2_IMAGE_SUBVOL_OBJECTID);
2396         if (!ext2_root) {
2397                 fprintf(stderr, "unable to create subvol\n");
2398                 goto fail;
2399         }
2400         ret = create_ext2_image(ext2_root, ext2_fs, "image", datacsum);
2401         if (ret) {
2402                 fprintf(stderr, "error during create_ext2_image %d\n", ret);
2403                 goto fail;
2404         }
2405         memset(root->fs_info->super_copy->label, 0, BTRFS_LABEL_SIZE);
2406         if (copylabel == 1) {
2407                 strncpy(root->fs_info->super_copy->label,
2408                                 ext2_fs->super->s_volume_name, 16);
2409                 fprintf(stderr, "copy label '%s'\n",
2410                                 root->fs_info->super_copy->label);
2411         } else if (copylabel == -1) {
2412                 strncpy(root->fs_info->super_copy->label, fslabel, BTRFS_LABEL_SIZE);
2413                 fprintf(stderr, "set label to '%s'\n", fslabel);
2414         }
2415
2416         printf("cleaning up system chunk.\n");
2417         ret = cleanup_sys_chunk(root, ext2_root);
2418         if (ret) {
2419                 fprintf(stderr, "error during cleanup_sys_chunk %d\n", ret);
2420                 goto fail;
2421         }
2422         ret = close_ctree(root);
2423         if (ret) {
2424                 fprintf(stderr, "error during close_ctree %d\n", ret);
2425                 goto fail;
2426         }
2427         close_ext2fs(ext2_fs);
2428
2429         /*
2430          * If this step succeed, we get a mountable btrfs. Otherwise
2431          * the ext2fs is left unchanged.
2432          */
2433         ret = migrate_super_block(fd, super_bytenr, blocksize);
2434         if (ret) {
2435                 fprintf(stderr, "unable to migrate super block\n");
2436                 goto fail;
2437         }
2438
2439         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2440         if (!root) {
2441                 fprintf(stderr, "unable to open ctree\n");
2442                 goto fail;
2443         }
2444         /* move chunk tree into system chunk. */
2445         ret = fixup_chunk_mapping(root);
2446         if (ret) {
2447                 fprintf(stderr, "error during fixup_chunk_tree\n");
2448                 goto fail;
2449         }
2450         ret = close_ctree(root);
2451         close(fd);
2452
2453         printf("conversion complete.\n");
2454         return 0;
2455 fail:
2456         if (fd != -1)
2457                 close(fd);
2458         fprintf(stderr, "conversion aborted.\n");
2459         return -1;
2460 }
2461
2462 static int may_rollback(struct btrfs_root *root)
2463 {
2464         struct btrfs_fs_info *info = root->fs_info;
2465         struct btrfs_multi_bio *multi = NULL;
2466         u64 bytenr;
2467         u64 length;
2468         u64 physical;
2469         u64 total_bytes;
2470         int num_stripes;
2471         int ret;
2472
2473         if (btrfs_super_num_devices(info->super_copy) != 1)
2474                 goto fail;
2475
2476         bytenr = BTRFS_SUPER_INFO_OFFSET;
2477         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
2478
2479         while (1) {
2480                 ret = btrfs_map_block(&info->mapping_tree, WRITE, bytenr,
2481                                       &length, &multi, 0, NULL);
2482                 if (ret) {
2483                         if (ret == -ENOENT) {
2484                                 /* removed block group at the tail */
2485                                 if (length == (u64)-1)
2486                                         break;
2487
2488                                 /* removed block group in the middle */
2489                                 goto next;
2490                         }
2491                         goto fail;
2492                 }
2493
2494                 num_stripes = multi->num_stripes;
2495                 physical = multi->stripes[0].physical;
2496                 kfree(multi);
2497
2498                 if (num_stripes != 1 || physical != bytenr)
2499                         goto fail;
2500 next:
2501                 bytenr += length;
2502                 if (bytenr >= total_bytes)
2503                         break;
2504         }
2505         return 0;
2506 fail:
2507         return -1;
2508 }
2509
2510 static int do_rollback(const char *devname)
2511 {
2512         int fd = -1;
2513         int ret;
2514         int i;
2515         struct btrfs_root *root;
2516         struct btrfs_root *ext2_root;
2517         struct btrfs_root *chunk_root;
2518         struct btrfs_dir_item *dir;
2519         struct btrfs_inode_item *inode;
2520         struct btrfs_file_extent_item *fi;
2521         struct btrfs_trans_handle *trans;
2522         struct extent_buffer *leaf;
2523         struct btrfs_block_group_cache *cache1;
2524         struct btrfs_block_group_cache *cache2;
2525         struct btrfs_key key;
2526         struct btrfs_path path;
2527         struct extent_io_tree io_tree;
2528         char *buf = NULL;
2529         char *name;
2530         u64 bytenr;
2531         u64 num_bytes;
2532         u64 root_dir;
2533         u64 objectid;
2534         u64 offset;
2535         u64 start;
2536         u64 end;
2537         u64 sb_bytenr;
2538         u64 first_free;
2539         u64 total_bytes;
2540         u32 sectorsize;
2541
2542         extent_io_tree_init(&io_tree);
2543
2544         fd = open(devname, O_RDWR);
2545         if (fd < 0) {
2546                 fprintf(stderr, "unable to open %s\n", devname);
2547                 goto fail;
2548         }
2549         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2550         if (!root) {
2551                 fprintf(stderr, "unable to open ctree\n");
2552                 goto fail;
2553         }
2554         ret = may_rollback(root);
2555         if (ret < 0) {
2556                 fprintf(stderr, "unable to do rollback\n");
2557                 goto fail;
2558         }
2559
2560         sectorsize = root->sectorsize;
2561         buf = malloc(sectorsize);
2562         if (!buf) {
2563                 fprintf(stderr, "unable to allocate memory\n");
2564                 goto fail;
2565         }
2566
2567         btrfs_init_path(&path);
2568
2569         key.objectid = EXT2_IMAGE_SUBVOL_OBJECTID;
2570         key.type = BTRFS_ROOT_ITEM_KEY;
2571         key.offset = (u64)-1;
2572         ext2_root = btrfs_read_fs_root(root->fs_info, &key);
2573         if (!ext2_root || IS_ERR(ext2_root)) {
2574                 fprintf(stderr, "unable to open subvol %llu\n",
2575                         (unsigned long long)key.objectid);
2576                 goto fail;
2577         }
2578
2579         name = "image";
2580         root_dir = btrfs_root_dirid(&root->root_item);
2581         dir = btrfs_lookup_dir_item(NULL, ext2_root, &path,
2582                                    root_dir, name, strlen(name), 0);
2583         if (!dir || IS_ERR(dir)) {
2584                 fprintf(stderr, "unable to find file %s\n", name);
2585                 goto fail;
2586         }
2587         leaf = path.nodes[0];
2588         btrfs_dir_item_key_to_cpu(leaf, dir, &key);
2589         btrfs_release_path(&path);
2590
2591         objectid = key.objectid;
2592
2593         ret = btrfs_lookup_inode(NULL, ext2_root, &path, &key, 0);
2594         if (ret) {
2595                 fprintf(stderr, "unable to find inode item\n");
2596                 goto fail;
2597         }
2598         leaf = path.nodes[0];
2599         inode = btrfs_item_ptr(leaf, path.slots[0], struct btrfs_inode_item);
2600         total_bytes = btrfs_inode_size(leaf, inode);
2601         btrfs_release_path(&path);
2602
2603         key.objectid = objectid;
2604         key.offset = 0;
2605         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
2606         ret = btrfs_search_slot(NULL, ext2_root, &key, &path, 0, 0);
2607         if (ret != 0) {
2608                 fprintf(stderr, "unable to find first file extent\n");
2609                 btrfs_release_path(&path);
2610                 goto fail;
2611         }
2612
2613         /* build mapping tree for the relocated blocks */
2614         for (offset = 0; offset < total_bytes; ) {
2615                 leaf = path.nodes[0];
2616                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2617                         ret = btrfs_next_leaf(root, &path);
2618                         if (ret != 0)
2619                                 break;
2620                         continue;
2621                 }
2622
2623                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2624                 if (key.objectid != objectid || key.offset != offset ||
2625                     btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2626                         break;
2627
2628                 fi = btrfs_item_ptr(leaf, path.slots[0],
2629                                     struct btrfs_file_extent_item);
2630                 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2631                         break;
2632                 if (btrfs_file_extent_compression(leaf, fi) ||
2633                     btrfs_file_extent_encryption(leaf, fi) ||
2634                     btrfs_file_extent_other_encoding(leaf, fi))
2635                         break;
2636
2637                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2638                 /* skip holes and direct mapped extents */
2639                 if (bytenr == 0 || bytenr == offset)
2640                         goto next_extent;
2641
2642                 bytenr += btrfs_file_extent_offset(leaf, fi);
2643                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
2644
2645                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2646                 cache2 =  btrfs_lookup_block_group(root->fs_info,
2647                                                    offset + num_bytes - 1);
2648                 if (!cache1 || cache1 != cache2 ||
2649                     (!(cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
2650                      !intersect_with_sb(offset, num_bytes)))
2651                         break;
2652
2653                 set_extent_bits(&io_tree, offset, offset + num_bytes - 1,
2654                                 EXTENT_LOCKED, GFP_NOFS);
2655                 set_state_private(&io_tree, offset, bytenr);
2656 next_extent:
2657                 offset += btrfs_file_extent_num_bytes(leaf, fi);
2658                 path.slots[0]++;
2659         }
2660         btrfs_release_path(&path);
2661
2662         if (offset < total_bytes) {
2663                 fprintf(stderr, "unable to build extent mapping\n");
2664                 goto fail;
2665         }
2666
2667         first_free = BTRFS_SUPER_INFO_OFFSET + 2 * sectorsize - 1;
2668         first_free &= ~((u64)sectorsize - 1);
2669         /* backup for extent #0 should exist */
2670         if(!test_range_bit(&io_tree, 0, first_free - 1, EXTENT_LOCKED, 1)) {
2671                 fprintf(stderr, "no backup for the first extent\n");
2672                 goto fail;
2673         }
2674         /* force no allocation from system block group */
2675         root->fs_info->system_allocs = -1;
2676         trans = btrfs_start_transaction(root, 1);
2677         BUG_ON(!trans);
2678         /*
2679          * recow the whole chunk tree, this will remove all chunk tree blocks
2680          * from system block group
2681          */
2682         chunk_root = root->fs_info->chunk_root;
2683         memset(&key, 0, sizeof(key));
2684         while (1) {
2685                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2686                 if (ret < 0)
2687                         break;
2688
2689                 ret = btrfs_next_leaf(chunk_root, &path);
2690                 if (ret)
2691                         break;
2692
2693                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2694                 btrfs_release_path(&path);
2695         }
2696         btrfs_release_path(&path);
2697
2698         offset = 0;
2699         num_bytes = 0;
2700         while(1) {
2701                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2702                 if (!cache1)
2703                         break;
2704
2705                 if (cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2706                         num_bytes += btrfs_block_group_used(&cache1->item);
2707
2708                 offset = cache1->key.objectid + cache1->key.offset;
2709         }
2710         /* only extent #0 left in system block group? */
2711         if (num_bytes > first_free) {
2712                 fprintf(stderr, "unable to empty system block group\n");
2713                 goto fail;
2714         }
2715         /* create a system chunk that maps the whole device */
2716         ret = prepare_system_chunk_sb(root->fs_info->super_copy);
2717         if (ret) {
2718                 fprintf(stderr, "unable to update system chunk\n");
2719                 goto fail;
2720         }
2721
2722         ret = btrfs_commit_transaction(trans, root);
2723         BUG_ON(ret);
2724
2725         ret = close_ctree(root);
2726         if (ret) {
2727                 fprintf(stderr, "error during close_ctree %d\n", ret);
2728                 goto fail;
2729         }
2730
2731         /* zero btrfs super block mirrors */
2732         memset(buf, 0, sectorsize);
2733         for (i = 1 ; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2734                 bytenr = btrfs_sb_offset(i);
2735                 if (bytenr >= total_bytes)
2736                         break;
2737                 ret = pwrite(fd, buf, sectorsize, bytenr);
2738                 if (ret != sectorsize) {
2739                         fprintf(stderr,
2740                                 "error during zeroing supreblock %d: %d\n",
2741                                 i, ret);
2742                         goto fail;
2743                 }
2744         }
2745
2746         sb_bytenr = (u64)-1;
2747         /* copy all relocated blocks back */
2748         while(1) {
2749                 ret = find_first_extent_bit(&io_tree, 0, &start, &end,
2750                                             EXTENT_LOCKED);
2751                 if (ret)
2752                         break;
2753
2754                 ret = get_state_private(&io_tree, start, &bytenr);
2755                 BUG_ON(ret);
2756
2757                 clear_extent_bits(&io_tree, start, end, EXTENT_LOCKED,
2758                                   GFP_NOFS);
2759
2760                 while (start <= end) {
2761                         if (start == BTRFS_SUPER_INFO_OFFSET) {
2762                                 sb_bytenr = bytenr;
2763                                 goto next_sector;
2764                         }
2765                         ret = pread(fd, buf, sectorsize, bytenr);
2766                         if (ret < 0) {
2767                                 fprintf(stderr, "error during pread %d\n", ret);
2768                                 goto fail;
2769                         }
2770                         BUG_ON(ret != sectorsize);
2771                         ret = pwrite(fd, buf, sectorsize, start);
2772                         if (ret < 0) {
2773                                 fprintf(stderr, "error during pwrite %d\n", ret);
2774                                 goto fail;
2775                         }
2776                         BUG_ON(ret != sectorsize);
2777 next_sector:
2778                         start += sectorsize;
2779                         bytenr += sectorsize;
2780                 }
2781         }
2782
2783         ret = fsync(fd);
2784         if (ret) {
2785                 fprintf(stderr, "error during fsync %d\n", ret);
2786                 goto fail;
2787         }
2788         /*
2789          * finally, overwrite btrfs super block.
2790          */
2791         ret = pread(fd, buf, sectorsize, sb_bytenr);
2792         if (ret < 0) {
2793                 fprintf(stderr, "error during pread %d\n", ret);
2794                 goto fail;
2795         }
2796         BUG_ON(ret != sectorsize);
2797         ret = pwrite(fd, buf, sectorsize, BTRFS_SUPER_INFO_OFFSET);
2798         if (ret < 0) {
2799                 fprintf(stderr, "error during pwrite %d\n", ret);
2800                 goto fail;
2801         }
2802         BUG_ON(ret != sectorsize);
2803         ret = fsync(fd);
2804         if (ret) {
2805                 fprintf(stderr, "error during fsync %d\n", ret);
2806                 goto fail;
2807         }
2808
2809         close(fd);
2810         free(buf);
2811         extent_io_tree_cleanup(&io_tree);
2812         printf("rollback complete.\n");
2813         return 0;
2814
2815 fail:
2816         if (fd != -1)
2817                 close(fd);
2818         free(buf);
2819         fprintf(stderr, "rollback aborted.\n");
2820         return -1;
2821 }
2822
2823 static void print_usage(void)
2824 {
2825         printf("usage: btrfs-convert [options] device\n");
2826         printf("options:\n");
2827         printf("\t-d|--no-datasum        disable data checksum, sets NODATASUM\n");
2828         printf("\t-i|--no-xattr          ignore xattrs and ACLs\n");
2829         printf("\t-n|--no-inline         disable inlining of small files to metadata\n");
2830         printf("\t-N|--nodesize SIZE     set filesystem metadata nodesize\n");
2831         printf("\t-r|--rollback          roll back to ext2fs\n");
2832         printf("\t-l|--label LABEL       set filesystem label\n");
2833         printf("\t-L|--copy-label        use label from converted filesystem\n");
2834         printf("\t-p|--progress          show converting progress (default)\n");
2835         printf("\t--no-progress          show only overview, not the detailed progress\n");
2836 }
2837
2838 int main(int argc, char *argv[])
2839 {
2840         int ret;
2841         int packing = 1;
2842         int noxattr = 0;
2843         int datacsum = 1;
2844         u32 nodesize = max_t(u32, sysconf(_SC_PAGESIZE),
2845                         BTRFS_MKFS_DEFAULT_NODE_SIZE);
2846         int rollback = 0;
2847         int copylabel = 0;
2848         int usage_error = 0;
2849         int progress = 1;
2850         char *file;
2851         char *fslabel = NULL;
2852
2853         while(1) {
2854                 enum { GETOPT_VAL_NO_PROGRESS = 256 };
2855                 static const struct option long_options[] = {
2856                         { "no-progress", no_argument, NULL,
2857                                 GETOPT_VAL_NO_PROGRESS },
2858                         { "no-datasum", no_argument, NULL, 'd' },
2859                         { "no-inline", no_argument, NULL, 'n' },
2860                         { "no-xattr", no_argument, NULL, 'i' },
2861                         { "rollback", no_argument, NULL, 'r' },
2862                         { "progress", no_argument, NULL, 'p' },
2863                         { "label", required_argument, NULL, 'l' },
2864                         { "copy-label", no_argument, NULL, 'L' },
2865                         { "nodesize", required_argument, NULL, 'N' },
2866                         { NULL, 0, NULL, 0 }
2867                 };
2868                 int c = getopt_long(argc, argv, "dinN:rl:Lp", long_options, NULL);
2869
2870                 if (c < 0)
2871                         break;
2872                 switch(c) {
2873                         case 'd':
2874                                 datacsum = 0;
2875                                 break;
2876                         case 'i':
2877                                 noxattr = 1;
2878                                 break;
2879                         case 'n':
2880                                 packing = 0;
2881                                 break;
2882                         case 'N':
2883                                 nodesize = parse_size(optarg);
2884                                 break;
2885                         case 'r':
2886                                 rollback = 1;
2887                                 break;
2888                         case 'l':
2889                                 copylabel = -1;
2890                                 fslabel = strdup(optarg);
2891                                 if (strlen(fslabel) > BTRFS_LABEL_SIZE) {
2892                                         fprintf(stderr,
2893                                                 "warning: label too long, trimmed to %d bytes\n",
2894                                                 BTRFS_LABEL_SIZE);
2895                                         fslabel[BTRFS_LABEL_SIZE] = 0;
2896                                 }
2897                                 break;
2898                         case 'L':
2899                                 copylabel = 1;
2900                                 break;
2901                         case 'p':
2902                                 progress = 1;
2903                                 break;
2904                         case GETOPT_VAL_NO_PROGRESS:
2905                                 progress = 0;
2906                                 break;
2907                         default:
2908                                 print_usage();
2909                                 return 1;
2910                 }
2911         }
2912         argc = argc - optind;
2913         set_argv0(argv);
2914         if (check_argc_exact(argc, 1)) {
2915                 print_usage();
2916                 return 1;
2917         }
2918
2919         if (rollback && (!datacsum || noxattr || !packing)) {
2920                 fprintf(stderr,
2921                         "Usage error: -d, -i, -n options do not apply to rollback\n");
2922                 usage_error++;
2923         }
2924
2925         if (usage_error) {
2926                 print_usage();
2927                 return 1;
2928         }
2929
2930         file = argv[optind];
2931         ret = check_mounted(file);
2932         if (ret < 0) {
2933                 fprintf(stderr, "Could not check mount status: %s\n",
2934                         strerror(-ret));
2935                 return 1;
2936         } else if (ret) {
2937                 fprintf(stderr, "%s is mounted\n", file);
2938                 return 1;
2939         }
2940
2941         if (rollback) {
2942                 ret = do_rollback(file);
2943         } else {
2944                 ret = do_convert(file, datacsum, packing, noxattr, nodesize,
2945                                 copylabel, fslabel, progress);
2946         }
2947         if (ret)
2948                 return 1;
2949         return 0;
2950 }