btrfs-convert.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include "kerncompat.h"
  20
  21 #include <sys/ioctl.h>
  22 #include <sys/mount.h>
  23 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <sys/types.h>
  26 #include <sys/stat.h>
  27 #include <sys/acl.h>
  28 #include <fcntl.h>
  29 #include <unistd.h>
  30 #include <uuid/uuid.h>
  31 #include <linux/limits.h>
  32 #include <getopt.h>
  33
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "volumes.h"
  37 #include "transaction.h"
  38 #include "crc32c.h"
  39 #include "utils.h"
  40 #include "task-utils.h"
  41 #include <ext2fs/ext2_fs.h>
  42 #include <ext2fs/ext2fs.h>
  43 #include <ext2fs/ext2_ext_attr.h>
  44
  45 #define INO_OFFSET (BTRFS_FIRST_FREE_OBJECTID - EXT2_ROOT_INO)
  46 #define EXT2_IMAGE_SUBVOL_OBJECTID BTRFS_FIRST_FREE_OBJECTID
  47
  48 struct task_ctx {
  49         uint32_t max_copy_inodes;
  50         uint32_t cur_copy_inodes;
  51         struct task_info *info;
  52 };
  53
  54 static void *print_copied_inodes(void *p)
  55 {
  56         struct task_ctx *priv = p;
  57         const char work_indicator[] = { '.', 'o', 'O', 'o' };
  58         uint32_t count = 0;
  59
  60         task_period_start(priv->info, 1000 /* 1s */);
  61         while (1) {
  62                 count++;
  63                 printf("copy inodes [%c] [%10d/%10d]\r",
  64                        work_indicator[count % 4], priv->cur_copy_inodes,
  65                        priv->max_copy_inodes);
  66                 fflush(stdout);
  67                 task_period_wait(priv->info);
  68         }
  69
  70         return NULL;
  71 }
  72
  73 static int after_copied_inodes(void *p)
  74 {
  75         struct task_ctx *priv = p;
  76
  77         printf("\n");
  78         task_period_stop(priv->info);
  79
  80         return 0;
  81 }
  82
  83 /*
  84  * Open Ext2fs in readonly mode, read block allocation bitmap and
  85  * inode bitmap into memory.
  86  */
  87 static int open_ext2fs(const char *name, ext2_filsys *ret_fs)
  88 {
  89         errcode_t ret;
  90         ext2_filsys ext2_fs;
  91         ext2_ino_t ino;
  92         ret = ext2fs_open(name, 0, 0, 0, unix_io_manager, &ext2_fs);
  93         if (ret) {
  94                 fprintf(stderr, "ext2fs_open: %s\n", error_message(ret));
  95                 goto fail;
  96         }
  97         ret = ext2fs_read_inode_bitmap(ext2_fs);
  98         if (ret) {
  99                 fprintf(stderr, "ext2fs_read_inode_bitmap: %s\n",
 100                         error_message(ret));
 101                 goto fail;
 102         }
 103         ret = ext2fs_read_block_bitmap(ext2_fs);
 104         if (ret) {
 105                 fprintf(stderr, "ext2fs_read_block_bitmap: %s\n",
 106                         error_message(ret));
 107                 goto fail;
 108         }
 109         /*
 110          * search each block group for a free inode. this set up
 111          * uninit block/inode bitmaps appropriately.
 112          */
 113         ino = 1;
 114         while (ino <= ext2_fs->super->s_inodes_count) {
 115                 ext2_ino_t foo;
 116                 ext2fs_new_inode(ext2_fs, ino, 0, NULL, &foo);
 117                 ino += EXT2_INODES_PER_GROUP(ext2_fs->super);
 118         }
 119
 120         *ret_fs = ext2_fs;
 121         return 0;
 122 fail:
 123         return -1;
 124 }
 125
 126 static int close_ext2fs(ext2_filsys fs)
 127 {
 128         ext2fs_close(fs);
 129         return 0;
 130 }
 131
 132 static int ext2_alloc_block(ext2_filsys fs, u64 goal, u64 *block_ret)
 133 {
 134         blk_t block;
 135
 136         if (!ext2fs_new_block(fs, goal, NULL, &block)) {
 137                 ext2fs_fast_mark_block_bitmap(fs->block_map, block);
 138                 *block_ret = block;
 139                 return 0;
 140         }
 141         return -ENOSPC;
 142 }
 143
 144 static int ext2_free_block(ext2_filsys fs, u64 block)
 145 {
 146         BUG_ON(block != (blk_t)block);
 147         ext2fs_fast_unmark_block_bitmap(fs->block_map, block);
 148         return 0;
 149 }
 150
 151 static int cache_free_extents(struct btrfs_root *root, ext2_filsys ext2_fs)
 152
 153 {
 154         int i, ret = 0;
 155         blk_t block;
 156         u64 bytenr;
 157         u64 blocksize = ext2_fs->blocksize;
 158
 159         block = ext2_fs->super->s_first_data_block;
 160         for (; block < ext2_fs->super->s_blocks_count; block++) {
 161                 if (ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
 162                         continue;
 163                 bytenr = block * blocksize;
 164                 ret = set_extent_dirty(&root->fs_info->free_space_cache,
 165                                        bytenr, bytenr + blocksize - 1, 0);
 166                 BUG_ON(ret);
 167         }
 168
 169         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 170                 bytenr = btrfs_sb_offset(i);
 171                 bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 172                 if (bytenr >= blocksize * ext2_fs->super->s_blocks_count)
 173                         break;
 174                 clear_extent_dirty(&root->fs_info->free_space_cache, bytenr,
 175                                    bytenr + BTRFS_STRIPE_LEN - 1, 0);
 176         }
 177
 178         clear_extent_dirty(&root->fs_info->free_space_cache,
 179                            0, BTRFS_SUPER_INFO_OFFSET - 1, 0);
 180
 181         return 0;
 182 }
 183
 184 static int custom_alloc_extent(struct btrfs_root *root, u64 num_bytes,
 185                                u64 hint_byte, struct btrfs_key *ins)
 186 {
 187         u64 start;
 188         u64 end;
 189         u64 last = hint_byte;
 190         int ret;
 191         int wrapped = 0;
 192         struct btrfs_block_group_cache *cache;
 193
 194         while(1) {
 195                 ret = find_first_extent_bit(&root->fs_info->free_space_cache,
 196                                             last, &start, &end, EXTENT_DIRTY);
 197                 if (ret) {
 198                         if (wrapped++ == 0) {
 199                                 last = 0;
 200                                 continue;
 201                         } else {
 202                                 goto fail;
 203                         }
 204                 }
 205
 206                 start = max(last, start);
 207                 last = end + 1;
 208                 if (last - start < num_bytes)
 209                         continue;
 210
 211                 last = start + num_bytes;
 212                 if (test_range_bit(&root->fs_info->pinned_extents,
 213                                    start, last - 1, EXTENT_DIRTY, 0))
 214                         continue;
 215
 216                 cache = btrfs_lookup_block_group(root->fs_info, start);
 217                 BUG_ON(!cache);
 218                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM ||
 219                     last > cache->key.objectid + cache->key.offset) {
 220                         last = cache->key.objectid + cache->key.offset;
 221                         continue;
 222                 }
 223
 224                 clear_extent_dirty(&root->fs_info->free_space_cache,
 225                                    start, start + num_bytes - 1, 0);
 226
 227                 ins->objectid = start;
 228                 ins->offset = num_bytes;
 229                 ins->type = BTRFS_EXTENT_ITEM_KEY;
 230                 return 0;
 231         }
 232 fail:
 233         fprintf(stderr, "not enough free space\n");
 234         return -ENOSPC;
 235 }
 236
 237 static int intersect_with_sb(u64 bytenr, u64 num_bytes)
 238 {
 239         int i;
 240         u64 offset;
 241
 242         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 243                 offset = btrfs_sb_offset(i);
 244                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
 245
 246                 if (bytenr < offset + BTRFS_STRIPE_LEN &&
 247                     bytenr + num_bytes > offset)
 248                         return 1;
 249         }
 250         return 0;
 251 }
 252
 253 static int custom_free_extent(struct btrfs_root *root, u64 bytenr,
 254                               u64 num_bytes)
 255 {
 256         return intersect_with_sb(bytenr, num_bytes);
 257 }
 258
 259 static struct btrfs_extent_ops extent_ops = {
 260         .alloc_extent = custom_alloc_extent,
 261         .free_extent = custom_free_extent,
 262 };
 263
 264 struct dir_iterate_data {
 265         struct btrfs_trans_handle *trans;
 266         struct btrfs_root *root;
 267         struct btrfs_inode_item *inode;
 268         u64 objectid;
 269         u64 index_cnt;
 270         u64 parent;
 271         int errcode;
 272 };
 273
 274 static u8 filetype_conversion_table[EXT2_FT_MAX] = {
 275         [EXT2_FT_UNKNOWN]       = BTRFS_FT_UNKNOWN,
 276         [EXT2_FT_REG_FILE]      = BTRFS_FT_REG_FILE,
 277         [EXT2_FT_DIR]           = BTRFS_FT_DIR,
 278         [EXT2_FT_CHRDEV]        = BTRFS_FT_CHRDEV,
 279         [EXT2_FT_BLKDEV]        = BTRFS_FT_BLKDEV,
 280         [EXT2_FT_FIFO]          = BTRFS_FT_FIFO,
 281         [EXT2_FT_SOCK]          = BTRFS_FT_SOCK,
 282         [EXT2_FT_SYMLINK]       = BTRFS_FT_SYMLINK,
 283 };
 284
 285 static int dir_iterate_proc(ext2_ino_t dir, int entry,
 286                             struct ext2_dir_entry *dirent,
 287                             int offset, int blocksize,
 288                             char *buf,void *priv_data)
 289 {
 290         int ret;
 291         int file_type;
 292         u64 objectid;
 293         u64 inode_size;
 294         char dotdot[] = "..";
 295         struct btrfs_key location;
 296         struct dir_iterate_data *idata = (struct dir_iterate_data *)priv_data;
 297         int name_len;
 298
 299         name_len = dirent->name_len & 0xFF;
 300
 301         objectid = dirent->inode + INO_OFFSET;
 302         if (!strncmp(dirent->name, dotdot, name_len)) {
 303                 if (name_len == 2) {
 304                         BUG_ON(idata->parent != 0);
 305                         idata->parent = objectid;
 306                 }
 307                 return 0;
 308         }
 309         if (dirent->inode < EXT2_GOOD_OLD_FIRST_INO)
 310                 return 0;
 311
 312         location.objectid = objectid;
 313         location.offset = 0;
 314         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
 315
 316         file_type = dirent->name_len >> 8;
 317         BUG_ON(file_type > EXT2_FT_SYMLINK);
 318         ret = btrfs_insert_dir_item(idata->trans, idata->root,
 319                                     dirent->name, name_len,
 320                                     idata->objectid, &location,
 321                                     filetype_conversion_table[file_type],
 322                                     idata->index_cnt);
 323         if (ret)
 324                 goto fail;
 325         ret = btrfs_insert_inode_ref(idata->trans, idata->root,
 326                                      dirent->name, name_len,
 327                                      objectid, idata->objectid,
 328                                      idata->index_cnt);
 329         if (ret)
 330                 goto fail;
 331         idata->index_cnt++;
 332         inode_size = btrfs_stack_inode_size(idata->inode) +
 333                      name_len * 2;
 334         btrfs_set_stack_inode_size(idata->inode, inode_size);
 335         return 0;
 336 fail:
 337         idata->errcode = ret;
 338         return BLOCK_ABORT;
 339 }
 340
 341 static int create_dir_entries(struct btrfs_trans_handle *trans,
 342                               struct btrfs_root *root, u64 objectid,
 343                               struct btrfs_inode_item *btrfs_inode,
 344                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 345 {
 346         int ret;
 347         errcode_t err;
 348         struct dir_iterate_data data = {
 349                 .trans          = trans,
 350                 .root           = root,
 351                 .inode          = btrfs_inode,
 352                 .objectid       = objectid,
 353                 .index_cnt      = 2,
 354                 .parent         = 0,
 355                 .errcode        = 0,
 356         };
 357
 358         err = ext2fs_dir_iterate2(ext2_fs, ext2_ino, 0, NULL,
 359                                   dir_iterate_proc, &data);
 360         if (err)
 361                 goto error;
 362         ret = data.errcode;
 363         if (ret == 0 && data.parent == objectid) {
 364                 ret = btrfs_insert_inode_ref(trans, root, "..", 2,
 365                                              objectid, objectid, 0);
 366         }
 367         return ret;
 368 error:
 369         fprintf(stderr, "ext2fs_dir_iterate2: %s\n", error_message(err));
 370         return -1;
 371 }
 372
 373 static int read_disk_extent(struct btrfs_root *root, u64 bytenr,
 374                             u32 num_bytes, char *buffer)
 375 {
 376         int ret;
 377         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
 378
 379         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, bytenr);
 380         if (ret != num_bytes)
 381                 goto fail;
 382         ret = 0;
 383 fail:
 384         if (ret > 0)
 385                 ret = -1;
 386         return ret;
 387 }
 388
 389 static int csum_disk_extent(struct btrfs_trans_handle *trans,
 390                             struct btrfs_root *root,
 391                             u64 disk_bytenr, u64 num_bytes)
 392 {
 393         u32 blocksize = root->sectorsize;
 394         u64 offset;
 395         char *buffer;
 396         int ret = 0;
 397
 398         buffer = malloc(blocksize);
 399         if (!buffer)
 400                 return -ENOMEM;
 401         for (offset = 0; offset < num_bytes; offset += blocksize) {
 402                 ret = read_disk_extent(root, disk_bytenr + offset,
 403                                         blocksize, buffer);
 404                 if (ret)
 405                         break;
 406                 ret = btrfs_csum_file_block(trans,
 407                                             root->fs_info->csum_root,
 408                                             disk_bytenr + num_bytes,
 409                                             disk_bytenr + offset,
 410                                             buffer, blocksize);
 411                 if (ret)
 412                         break;
 413         }
 414         free(buffer);
 415         return ret;
 416 }
 417
 418 static int record_file_blocks(struct btrfs_trans_handle *trans,
 419                               struct btrfs_root *root, u64 objectid,
 420                               struct btrfs_inode_item *inode,
 421                               u64 file_block, u64 disk_block,
 422                               u64 num_blocks, int checksum)
 423 {
 424         int ret;
 425         u64 file_pos = file_block * root->sectorsize;
 426         u64 disk_bytenr = disk_block * root->sectorsize;
 427         u64 num_bytes = num_blocks * root->sectorsize;
 428         ret = btrfs_record_file_extent(trans, root, objectid, inode, file_pos,
 429                                         disk_bytenr, num_bytes);
 430
 431         if (ret || !checksum || disk_bytenr == 0)
 432                 return ret;
 433
 434         return csum_disk_extent(trans, root, disk_bytenr, num_bytes);
 435 }
 436
 437 struct blk_iterate_data {
 438         struct btrfs_trans_handle *trans;
 439         struct btrfs_root *root;
 440         struct btrfs_inode_item *inode;
 441         u64 objectid;
 442         u64 first_block;
 443         u64 disk_block;
 444         u64 num_blocks;
 445         u64 boundary;
 446         int checksum;
 447         int errcode;
 448 };
 449
 450 static int block_iterate_proc(ext2_filsys ext2_fs,
 451                               u64 disk_block, u64 file_block,
 452                               struct blk_iterate_data *idata)
 453 {
 454         int ret;
 455         int sb_region;
 456         int do_barrier;
 457         struct btrfs_root *root = idata->root;
 458         struct btrfs_trans_handle *trans = idata->trans;
 459         struct btrfs_block_group_cache *cache;
 460         u64 bytenr = disk_block * root->sectorsize;
 461
 462         sb_region = intersect_with_sb(bytenr, root->sectorsize);
 463         do_barrier = sb_region || disk_block >= idata->boundary;
 464         if ((idata->num_blocks > 0 && do_barrier) ||
 465             (file_block > idata->first_block + idata->num_blocks) ||
 466             (disk_block != idata->disk_block + idata->num_blocks)) {
 467                 if (idata->num_blocks > 0) {
 468                         ret = record_file_blocks(trans, root, idata->objectid,
 469                                         idata->inode, idata->first_block,
 470                                         idata->disk_block, idata->num_blocks,
 471                                         idata->checksum);
 472                         if (ret)
 473                                 goto fail;
 474                         idata->first_block += idata->num_blocks;
 475                         idata->num_blocks = 0;
 476                 }
 477                 if (file_block > idata->first_block) {
 478                         ret = record_file_blocks(trans, root, idata->objectid,
 479                                         idata->inode, idata->first_block,
 480                                         0, file_block - idata->first_block,
 481                                         idata->checksum);
 482                         if (ret)
 483                                 goto fail;
 484                 }
 485
 486                 if (sb_region) {
 487                         bytenr += BTRFS_STRIPE_LEN - 1;
 488                         bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 489                 } else {
 490                         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
 491                         BUG_ON(!cache);
 492                         bytenr = cache->key.objectid + cache->key.offset;
 493                 }
 494
 495                 idata->first_block = file_block;
 496                 idata->disk_block = disk_block;
 497                 idata->boundary = bytenr / root->sectorsize;
 498         }
 499         idata->num_blocks++;
 500         return 0;
 501 fail:
 502         idata->errcode = ret;
 503         return BLOCK_ABORT;
 504 }
 505
 506 static int __block_iterate_proc(ext2_filsys fs, blk_t *blocknr,
 507                                 e2_blkcnt_t blockcnt, blk_t ref_block,
 508                                 int ref_offset, void *priv_data)
 509 {
 510         struct blk_iterate_data *idata;
 511         idata = (struct blk_iterate_data *)priv_data;
 512         return block_iterate_proc(fs, *blocknr, blockcnt, idata);
 513 }
 514
 515 /*
 516  * traverse file's data blocks, record these data blocks as file extents.
 517  */
 518 static int create_file_extents(struct btrfs_trans_handle *trans,
 519                                struct btrfs_root *root, u64 objectid,
 520                                struct btrfs_inode_item *btrfs_inode,
 521                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 522                                int datacsum, int packing)
 523 {
 524         int ret;
 525         char *buffer = NULL;
 526         errcode_t err;
 527         u32 last_block;
 528         u32 sectorsize = root->sectorsize;
 529         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 530         struct blk_iterate_data data = {
 531                 .trans          = trans,
 532                 .root           = root,
 533                 .inode          = btrfs_inode,
 534                 .objectid       = objectid,
 535                 .first_block    = 0,
 536                 .disk_block     = 0,
 537                 .num_blocks     = 0,
 538                 .boundary       = (u64)-1,
 539                 .checksum       = datacsum,
 540                 .errcode        = 0,
 541         };
 542         err = ext2fs_block_iterate2(ext2_fs, ext2_ino, BLOCK_FLAG_DATA_ONLY,
 543                                     NULL, __block_iterate_proc, &data);
 544         if (err)
 545                 goto error;
 546         ret = data.errcode;
 547         if (ret)
 548                 goto fail;
 549         if (packing && data.first_block == 0 && data.num_blocks > 0 &&
 550             inode_size <= BTRFS_MAX_INLINE_DATA_SIZE(root)) {
 551                 u64 num_bytes = data.num_blocks * sectorsize;
 552                 u64 disk_bytenr = data.disk_block * sectorsize;
 553                 u64 nbytes;
 554
 555                 buffer = malloc(num_bytes);
 556                 if (!buffer)
 557                         return -ENOMEM;
 558                 ret = read_disk_extent(root, disk_bytenr, num_bytes, buffer);
 559                 if (ret)
 560                         goto fail;
 561                 if (num_bytes > inode_size)
 562                         num_bytes = inode_size;
 563                 ret = btrfs_insert_inline_extent(trans, root, objectid,
 564                                                  0, buffer, num_bytes);
 565                 if (ret)
 566                         goto fail;
 567                 nbytes = btrfs_stack_inode_nbytes(btrfs_inode) + num_bytes;
 568                 btrfs_set_stack_inode_nbytes(btrfs_inode, nbytes);
 569         } else if (data.num_blocks > 0) {
 570                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 571                                          data.first_block, data.disk_block,
 572                                          data.num_blocks, data.checksum);
 573                 if (ret)
 574                         goto fail;
 575         }
 576         data.first_block += data.num_blocks;
 577         last_block = (inode_size + sectorsize - 1) / sectorsize;
 578         if (last_block > data.first_block) {
 579                 ret = record_file_blocks(trans, root, objectid, btrfs_inode,
 580                                          data.first_block, 0, last_block -
 581                                          data.first_block, data.checksum);
 582         }
 583 fail:
 584         free(buffer);
 585         return ret;
 586 error:
 587         fprintf(stderr, "ext2fs_block_iterate2: %s\n", error_message(err));
 588         return -1;
 589 }
 590
 591 static int create_symbol_link(struct btrfs_trans_handle *trans,
 592                               struct btrfs_root *root, u64 objectid,
 593                               struct btrfs_inode_item *btrfs_inode,
 594                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 595                               struct ext2_inode *ext2_inode)
 596 {
 597         int ret;
 598         char *pathname;
 599         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 600         if (ext2fs_inode_data_blocks(ext2_fs, ext2_inode)) {
 601                 btrfs_set_stack_inode_size(btrfs_inode, inode_size + 1);
 602                 ret = create_file_extents(trans, root, objectid, btrfs_inode,
 603                                           ext2_fs, ext2_ino, 1, 1);
 604                 btrfs_set_stack_inode_size(btrfs_inode, inode_size);
 605                 return ret;
 606         }
 607
 608         pathname = (char *)&(ext2_inode->i_block[0]);
 609         BUG_ON(pathname[inode_size] != 0);
 610         ret = btrfs_insert_inline_extent(trans, root, objectid, 0,
 611                                          pathname, inode_size + 1);
 612         btrfs_set_stack_inode_nbytes(btrfs_inode, inode_size + 1);
 613         return ret;
 614 }
 615
 616 /*
 617  * Following xattr/acl related codes are based on codes in
 618  * fs/ext3/xattr.c and fs/ext3/acl.c
 619  */
 620 #define EXT2_XATTR_BHDR(ptr) ((struct ext2_ext_attr_header *)(ptr))
 621 #define EXT2_XATTR_BFIRST(ptr) \
 622         ((struct ext2_ext_attr_entry *)(EXT2_XATTR_BHDR(ptr) + 1))
 623 #define EXT2_XATTR_IHDR(inode) \
 624         ((struct ext2_ext_attr_header *) ((void *)(inode) + \
 625                 EXT2_GOOD_OLD_INODE_SIZE + (inode)->i_extra_isize))
 626 #define EXT2_XATTR_IFIRST(inode) \
 627         ((struct ext2_ext_attr_entry *) ((void *)EXT2_XATTR_IHDR(inode) + \
 628                 sizeof(EXT2_XATTR_IHDR(inode)->h_magic)))
 629
 630 static int ext2_xattr_check_names(struct ext2_ext_attr_entry *entry,
 631                                   const void *end)
 632 {
 633         struct ext2_ext_attr_entry *next;
 634
 635         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 636                 next = EXT2_EXT_ATTR_NEXT(entry);
 637                 if ((void *)next >= end)
 638                         return -EIO;
 639                 entry = next;
 640         }
 641         return 0;
 642 }
 643
 644 static int ext2_xattr_check_block(const char *buf, size_t size)
 645 {
 646         int error;
 647         struct ext2_ext_attr_header *header = EXT2_XATTR_BHDR(buf);
 648
 649         if (header->h_magic != EXT2_EXT_ATTR_MAGIC ||
 650             header->h_blocks != 1)
 651                 return -EIO;
 652         error = ext2_xattr_check_names(EXT2_XATTR_BFIRST(buf), buf + size);
 653         return error;
 654 }
 655
 656 static int ext2_xattr_check_entry(struct ext2_ext_attr_entry *entry,
 657                                   size_t size)
 658 {
 659         size_t value_size = entry->e_value_size;
 660
 661         if (entry->e_value_block != 0 || value_size > size ||
 662             entry->e_value_offs + value_size > size)
 663                 return -EIO;
 664         return 0;
 665 }
 666
 667 #define EXT2_ACL_VERSION        0x0001
 668
 669 typedef struct {
 670         __le16          e_tag;
 671         __le16          e_perm;
 672         __le32          e_id;
 673 } ext2_acl_entry;
 674
 675 typedef struct {
 676         __le16          e_tag;
 677         __le16          e_perm;
 678 } ext2_acl_entry_short;
 679
 680 typedef struct {
 681         __le32          a_version;
 682 } ext2_acl_header;
 683
 684 static inline int ext2_acl_count(size_t size)
 685 {
 686         ssize_t s;
 687         size -= sizeof(ext2_acl_header);
 688         s = size - 4 * sizeof(ext2_acl_entry_short);
 689         if (s < 0) {
 690                 if (size % sizeof(ext2_acl_entry_short))
 691                         return -1;
 692                 return size / sizeof(ext2_acl_entry_short);
 693         } else {
 694                 if (s % sizeof(ext2_acl_entry))
 695                         return -1;
 696                 return s / sizeof(ext2_acl_entry) + 4;
 697         }
 698 }
 699
 700 #define ACL_EA_VERSION          0x0002
 701
 702 typedef struct {
 703         __le16          e_tag;
 704         __le16          e_perm;
 705         __le32          e_id;
 706 } acl_ea_entry;
 707
 708 typedef struct {
 709         __le32          a_version;
 710         acl_ea_entry    a_entries[0];
 711 } acl_ea_header;
 712
 713 static inline size_t acl_ea_size(int count)
 714 {
 715         return sizeof(acl_ea_header) + count * sizeof(acl_ea_entry);
 716 }
 717
 718 static int ext2_acl_to_xattr(void *dst, const void *src,
 719                              size_t dst_size, size_t src_size)
 720 {
 721         int i, count;
 722         const void *end = src + src_size;
 723         acl_ea_header *ext_acl = (acl_ea_header *)dst;
 724         acl_ea_entry *dst_entry = ext_acl->a_entries;
 725         ext2_acl_entry *src_entry;
 726
 727         if (src_size < sizeof(ext2_acl_header))
 728                 goto fail;
 729         if (((ext2_acl_header *)src)->a_version !=
 730             cpu_to_le32(EXT2_ACL_VERSION))
 731                 goto fail;
 732         src += sizeof(ext2_acl_header);
 733         count = ext2_acl_count(src_size);
 734         if (count <= 0)
 735                 goto fail;
 736
 737         BUG_ON(dst_size < acl_ea_size(count));
 738         ext_acl->a_version = cpu_to_le32(ACL_EA_VERSION);
 739         for (i = 0; i < count; i++, dst_entry++) {
 740                 src_entry = (ext2_acl_entry *)src;
 741                 if (src + sizeof(ext2_acl_entry_short) > end)
 742                         goto fail;
 743                 dst_entry->e_tag = src_entry->e_tag;
 744                 dst_entry->e_perm = src_entry->e_perm;
 745                 switch (le16_to_cpu(src_entry->e_tag)) {
 746                 case ACL_USER_OBJ:
 747                 case ACL_GROUP_OBJ:
 748                 case ACL_MASK:
 749                 case ACL_OTHER:
 750                         src += sizeof(ext2_acl_entry_short);
 751                         dst_entry->e_id = cpu_to_le32(ACL_UNDEFINED_ID);
 752                         break;
 753                 case ACL_USER:
 754                 case ACL_GROUP:
 755                         src += sizeof(ext2_acl_entry);
 756                         if (src > end)
 757                                 goto fail;
 758                         dst_entry->e_id = src_entry->e_id;
 759                         break;
 760                 default:
 761                         goto fail;
 762                 }
 763         }
 764         if (src != end)
 765                 goto fail;
 766         return 0;
 767 fail:
 768         return -EINVAL;
 769 }
 770
 771 static char *xattr_prefix_table[] = {
 772         [1] =   "user.",
 773         [2] =   "system.posix_acl_access",
 774         [3] =   "system.posix_acl_default",
 775         [4] =   "trusted.",
 776         [6] =   "security.",
 777 };
 778
 779 static int copy_single_xattr(struct btrfs_trans_handle *trans,
 780                              struct btrfs_root *root, u64 objectid,
 781                              struct ext2_ext_attr_entry *entry,
 782                              const void *data, u32 datalen)
 783 {
 784         int ret = 0;
 785         int name_len;
 786         int name_index;
 787         void *databuf = NULL;
 788         char namebuf[XATTR_NAME_MAX + 1];
 789
 790         name_index = entry->e_name_index;
 791         if (name_index >= ARRAY_SIZE(xattr_prefix_table) ||
 792             xattr_prefix_table[name_index] == NULL)
 793                 return -EOPNOTSUPP;
 794         name_len = strlen(xattr_prefix_table[name_index]) +
 795                    entry->e_name_len;
 796         if (name_len >= sizeof(namebuf))
 797                 return -ERANGE;
 798
 799         if (name_index == 2 || name_index == 3) {
 800                 size_t bufsize = acl_ea_size(ext2_acl_count(datalen));
 801                 databuf = malloc(bufsize);
 802                 if (!databuf)
 803                        return -ENOMEM;
 804                 ret = ext2_acl_to_xattr(databuf, data, bufsize, datalen);
 805                 if (ret)
 806                         goto out;
 807                 data = databuf;
 808                 datalen = bufsize;
 809         }
 810         strncpy(namebuf, xattr_prefix_table[name_index], XATTR_NAME_MAX);
 811         strncat(namebuf, EXT2_EXT_ATTR_NAME(entry), entry->e_name_len);
 812         if (name_len + datalen > BTRFS_LEAF_DATA_SIZE(root) -
 813             sizeof(struct btrfs_item) - sizeof(struct btrfs_dir_item)) {
 814                 fprintf(stderr, "skip large xattr on inode %Lu name %.*s\n",
 815                         objectid - INO_OFFSET, name_len, namebuf);
 816                 goto out;
 817         }
 818         ret = btrfs_insert_xattr_item(trans, root, namebuf, name_len,
 819                                       data, datalen, objectid);
 820 out:
 821         free(databuf);
 822         return ret;
 823 }
 824
 825 static int copy_extended_attrs(struct btrfs_trans_handle *trans,
 826                                struct btrfs_root *root, u64 objectid,
 827                                struct btrfs_inode_item *btrfs_inode,
 828                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 829 {
 830         int ret = 0;
 831         int inline_ea = 0;
 832         errcode_t err;
 833         u32 datalen;
 834         u32 block_size = ext2_fs->blocksize;
 835         u32 inode_size = EXT2_INODE_SIZE(ext2_fs->super);
 836         struct ext2_inode_large *ext2_inode;
 837         struct ext2_ext_attr_entry *entry;
 838         void *data;
 839         char *buffer = NULL;
 840         char inode_buf[EXT2_GOOD_OLD_INODE_SIZE];
 841
 842         if (inode_size <= EXT2_GOOD_OLD_INODE_SIZE) {
 843                 ext2_inode = (struct ext2_inode_large *)inode_buf;
 844         } else {
 845                 ext2_inode = (struct ext2_inode_large *)malloc(inode_size);
 846                 if (!ext2_inode)
 847                        return -ENOMEM;
 848         }
 849         err = ext2fs_read_inode_full(ext2_fs, ext2_ino, (void *)ext2_inode,
 850                                      inode_size);
 851         if (err) {
 852                 fprintf(stderr, "ext2fs_read_inode_full: %s\n",
 853                         error_message(err));
 854                 ret = -1;
 855                 goto out;
 856         }
 857
 858         if (ext2_ino > ext2_fs->super->s_first_ino &&
 859             inode_size > EXT2_GOOD_OLD_INODE_SIZE) {
 860                 if (EXT2_GOOD_OLD_INODE_SIZE +
 861                     ext2_inode->i_extra_isize > inode_size) {
 862                         ret = -EIO;
 863                         goto out;
 864                 }
 865                 if (ext2_inode->i_extra_isize != 0 &&
 866                     EXT2_XATTR_IHDR(ext2_inode)->h_magic ==
 867                     EXT2_EXT_ATTR_MAGIC) {
 868                         inline_ea = 1;
 869                 }
 870         }
 871         if (inline_ea) {
 872                 int total;
 873                 void *end = (void *)ext2_inode + inode_size;
 874                 entry = EXT2_XATTR_IFIRST(ext2_inode);
 875                 total = end - (void *)entry;
 876                 ret = ext2_xattr_check_names(entry, end);
 877                 if (ret)
 878                         goto out;
 879                 while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 880                         ret = ext2_xattr_check_entry(entry, total);
 881                         if (ret)
 882                                 goto out;
 883                         data = (void *)EXT2_XATTR_IFIRST(ext2_inode) +
 884                                 entry->e_value_offs;
 885                         datalen = entry->e_value_size;
 886                         ret = copy_single_xattr(trans, root, objectid,
 887                                                 entry, data, datalen);
 888                         if (ret)
 889                                 goto out;
 890                         entry = EXT2_EXT_ATTR_NEXT(entry);
 891                 }
 892         }
 893
 894         if (ext2_inode->i_file_acl == 0)
 895                 goto out;
 896
 897         buffer = malloc(block_size);
 898         if (!buffer) {
 899                 ret = -ENOMEM;
 900                 goto out;
 901         }
 902         err = ext2fs_read_ext_attr(ext2_fs, ext2_inode->i_file_acl, buffer);
 903         if (err) {
 904                 fprintf(stderr, "ext2fs_read_ext_attr: %s\n",
 905                         error_message(err));
 906                 ret = -1;
 907                 goto out;
 908         }
 909         ret = ext2_xattr_check_block(buffer, block_size);
 910         if (ret)
 911                 goto out;
 912
 913         entry = EXT2_XATTR_BFIRST(buffer);
 914         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 915                 ret = ext2_xattr_check_entry(entry, block_size);
 916                 if (ret)
 917                         goto out;
 918                 data = buffer + entry->e_value_offs;
 919                 datalen = entry->e_value_size;
 920                 ret = copy_single_xattr(trans, root, objectid,
 921                                         entry, data, datalen);
 922                 if (ret)
 923                         goto out;
 924                 entry = EXT2_EXT_ATTR_NEXT(entry);
 925         }
 926 out:
 927         free(buffer);
 928         if ((void *)ext2_inode != inode_buf)
 929                 free(ext2_inode);
 930         return ret;
 931 }
 932 #define MINORBITS       20
 933 #define MKDEV(ma, mi)   (((ma) << MINORBITS) | (mi))
 934
 935 static inline dev_t old_decode_dev(u16 val)
 936 {
 937         return MKDEV((val >> 8) & 255, val & 255);
 938 }
 939
 940 static inline dev_t new_decode_dev(u32 dev)
 941 {
 942         unsigned major = (dev & 0xfff00) >> 8;
 943         unsigned minor = (dev & 0xff) | ((dev >> 12) & 0xfff00);
 944         return MKDEV(major, minor);
 945 }
 946
 947 static int copy_inode_item(struct btrfs_inode_item *dst,
 948                            struct ext2_inode *src, u32 blocksize)
 949 {
 950         btrfs_set_stack_inode_generation(dst, 1);
 951         btrfs_set_stack_inode_size(dst, src->i_size);
 952         btrfs_set_stack_inode_nbytes(dst, 0);
 953         btrfs_set_stack_inode_block_group(dst, 0);
 954         btrfs_set_stack_inode_nlink(dst, src->i_links_count);
 955         btrfs_set_stack_inode_uid(dst, src->i_uid | (src->i_uid_high << 16));
 956         btrfs_set_stack_inode_gid(dst, src->i_gid | (src->i_gid_high << 16));
 957         btrfs_set_stack_inode_mode(dst, src->i_mode);
 958         btrfs_set_stack_inode_rdev(dst, 0);
 959         btrfs_set_stack_inode_flags(dst, 0);
 960         btrfs_set_stack_timespec_sec(&dst->atime, src->i_atime);
 961         btrfs_set_stack_timespec_nsec(&dst->atime, 0);
 962         btrfs_set_stack_timespec_sec(&dst->ctime, src->i_ctime);
 963         btrfs_set_stack_timespec_nsec(&dst->ctime, 0);
 964         btrfs_set_stack_timespec_sec(&dst->mtime, src->i_mtime);
 965         btrfs_set_stack_timespec_nsec(&dst->mtime, 0);
 966         btrfs_set_stack_timespec_sec(&dst->otime, 0);
 967         btrfs_set_stack_timespec_nsec(&dst->otime, 0);
 968
 969         if (S_ISDIR(src->i_mode)) {
 970                 btrfs_set_stack_inode_size(dst, 0);
 971                 btrfs_set_stack_inode_nlink(dst, 1);
 972         }
 973         if (S_ISREG(src->i_mode)) {
 974                 btrfs_set_stack_inode_size(dst, (u64)src->i_size_high << 32 |
 975                                            (u64)src->i_size);
 976         }
 977         if (!S_ISREG(src->i_mode) && !S_ISDIR(src->i_mode) &&
 978             !S_ISLNK(src->i_mode)) {
 979                 if (src->i_block[0]) {
 980                         btrfs_set_stack_inode_rdev(dst,
 981                                 old_decode_dev(src->i_block[0]));
 982                 } else {
 983                         btrfs_set_stack_inode_rdev(dst,
 984                                 new_decode_dev(src->i_block[1]));
 985                 }
 986         }
 987         return 0;
 988 }
 989
 990 /*
 991  * copy a single inode. do all the required works, such as cloning
 992  * inode item, creating file extents and creating directory entries.
 993  */
 994 static int copy_single_inode(struct btrfs_trans_handle *trans,
 995                              struct btrfs_root *root, u64 objectid,
 996                              ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 997                              struct ext2_inode *ext2_inode,
 998                              int datacsum, int packing, int noxattr)
 999 {
1000         int ret;
1001         struct btrfs_key inode_key;
1002         struct btrfs_inode_item btrfs_inode;
1003
1004         if (ext2_inode->i_links_count == 0)
1005                 return 0;
1006
1007         copy_inode_item(&btrfs_inode, ext2_inode, ext2_fs->blocksize);
1008         if (!datacsum && S_ISREG(ext2_inode->i_mode)) {
1009                 u32 flags = btrfs_stack_inode_flags(&btrfs_inode) |
1010                             BTRFS_INODE_NODATASUM;
1011                 btrfs_set_stack_inode_flags(&btrfs_inode, flags);
1012         }
1013
1014         switch (ext2_inode->i_mode & S_IFMT) {
1015         case S_IFREG:
1016                 ret = create_file_extents(trans, root, objectid, &btrfs_inode,
1017                                         ext2_fs, ext2_ino, datacsum, packing);
1018                 break;
1019         case S_IFDIR:
1020                 ret = create_dir_entries(trans, root, objectid, &btrfs_inode,
1021                                          ext2_fs, ext2_ino);
1022                 break;
1023         case S_IFLNK:
1024                 ret = create_symbol_link(trans, root, objectid, &btrfs_inode,
1025                                          ext2_fs, ext2_ino, ext2_inode);
1026                 break;
1027         default:
1028                 ret = 0;
1029                 break;
1030         }
1031         if (ret)
1032                 return ret;
1033
1034         if (!noxattr) {
1035                 ret = copy_extended_attrs(trans, root, objectid, &btrfs_inode,
1036                                           ext2_fs, ext2_ino);
1037                 if (ret)
1038                         return ret;
1039         }
1040         inode_key.objectid = objectid;
1041         inode_key.offset = 0;
1042         btrfs_set_key_type(&inode_key, BTRFS_INODE_ITEM_KEY);
1043         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1044         return ret;
1045 }
1046
1047 static int copy_disk_extent(struct btrfs_root *root, u64 dst_bytenr,
1048                             u64 src_bytenr, u32 num_bytes)
1049 {
1050         int ret;
1051         char *buffer;
1052         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
1053
1054         buffer = malloc(num_bytes);
1055         if (!buffer)
1056                 return -ENOMEM;
1057         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, src_bytenr);
1058         if (ret != num_bytes)
1059                 goto fail;
1060         ret = pwrite(fs_devs->latest_bdev, buffer, num_bytes, dst_bytenr);
1061         if (ret != num_bytes)
1062                 goto fail;
1063         ret = 0;
1064 fail:
1065         free(buffer);
1066         if (ret > 0)
1067                 ret = -1;
1068         return ret;
1069 }
1070 /*
1071  * scan ext2's inode bitmap and copy all used inodes.
1072  */
1073 static int copy_inodes(struct btrfs_root *root, ext2_filsys ext2_fs,
1074                        int datacsum, int packing, int noxattr, struct task_ctx *p)
1075 {
1076         int ret;
1077         errcode_t err;
1078         ext2_inode_scan ext2_scan;
1079         struct ext2_inode ext2_inode;
1080         ext2_ino_t ext2_ino;
1081         u64 objectid;
1082         struct btrfs_trans_handle *trans;
1083
1084         trans = btrfs_start_transaction(root, 1);
1085         if (!trans)
1086                 return -ENOMEM;
1087         err = ext2fs_open_inode_scan(ext2_fs, 0, &ext2_scan);
1088         if (err) {
1089                 fprintf(stderr, "ext2fs_open_inode_scan: %s\n", error_message(err));
1090                 return -1;
1091         }
1092         while (!(err = ext2fs_get_next_inode(ext2_scan, &ext2_ino,
1093                                              &ext2_inode))) {
1094                 /* no more inodes */
1095                 if (ext2_ino == 0)
1096                         break;
1097                 /* skip special inode in ext2fs */
1098                 if (ext2_ino < EXT2_GOOD_OLD_FIRST_INO &&
1099                     ext2_ino != EXT2_ROOT_INO)
1100                         continue;
1101                 objectid = ext2_ino + INO_OFFSET;
1102                 ret = copy_single_inode(trans, root,
1103                                         objectid, ext2_fs, ext2_ino,
1104                                         &ext2_inode, datacsum, packing,
1105                                         noxattr);
1106                 p->cur_copy_inodes++;
1107                 if (ret)
1108                         return ret;
1109                 if (trans->blocks_used >= 4096) {
1110                         ret = btrfs_commit_transaction(trans, root);
1111                         BUG_ON(ret);
1112                         trans = btrfs_start_transaction(root, 1);
1113                         BUG_ON(!trans);
1114                 }
1115         }
1116         if (err) {
1117                 fprintf(stderr, "ext2fs_get_next_inode: %s\n", error_message(err));
1118                 return -1;
1119         }
1120         ret = btrfs_commit_transaction(trans, root);
1121         BUG_ON(ret);
1122
1123         return ret;
1124 }
1125
1126 /*
1127  * Construct a range of ext2fs image file.
1128  * scan block allocation bitmap, find all blocks used by the ext2fs
1129  * in this range and create file extents that point to these blocks.
1130  *
1131  * Note: Before calling the function, no file extent points to blocks
1132  *       in this range
1133  */
1134 static int create_image_file_range(struct btrfs_trans_handle *trans,
1135                                    struct btrfs_root *root, u64 objectid,
1136                                    struct btrfs_inode_item *inode,
1137                                    u64 start_byte, u64 end_byte,
1138                                    ext2_filsys ext2_fs)
1139 {
1140         u32 blocksize = ext2_fs->blocksize;
1141         u32 block = start_byte / blocksize;
1142         u32 last_block = (end_byte + blocksize - 1) / blocksize;
1143         int ret = 0;
1144         struct blk_iterate_data data = {
1145                 .trans          = trans,
1146                 .root           = root,
1147                 .inode          = inode,
1148                 .objectid       = objectid,
1149                 .first_block    = block,
1150                 .disk_block     = 0,
1151                 .num_blocks     = 0,
1152                 .boundary       = (u64)-1,
1153                 .checksum       = 0,
1154                 .errcode        = 0,
1155         };
1156         for (; start_byte < end_byte; block++, start_byte += blocksize) {
1157                 if (!ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
1158                         continue;
1159                 ret = block_iterate_proc(NULL, block, block, &data);
1160                 if (ret & BLOCK_ABORT) {
1161                         ret = data.errcode;
1162                         goto fail;
1163                 }
1164         }
1165         if (data.num_blocks > 0) {
1166                 ret = record_file_blocks(trans, root, objectid, inode,
1167                                          data.first_block, data.disk_block,
1168                                          data.num_blocks, 0);
1169                 if (ret)
1170                         goto fail;
1171                 data.first_block += data.num_blocks;
1172         }
1173         if (last_block > data.first_block) {
1174                 ret = record_file_blocks(trans, root, objectid, inode,
1175                                          data.first_block, 0, last_block -
1176                                          data.first_block, 0);
1177                 if (ret)
1178                         goto fail;
1179         }
1180 fail:
1181         return ret;
1182 }
1183 /*
1184  * Create the ext2fs image file.
1185  */
1186 static int create_ext2_image(struct btrfs_root *root, ext2_filsys ext2_fs,
1187                              const char *name)
1188 {
1189         int ret;
1190         struct btrfs_key key;
1191         struct btrfs_key location;
1192         struct btrfs_path path;
1193         struct btrfs_inode_item btrfs_inode;
1194         struct btrfs_inode_item *inode_item;
1195         struct extent_buffer *leaf;
1196         struct btrfs_fs_info *fs_info = root->fs_info;
1197         struct btrfs_root *extent_root = fs_info->extent_root;
1198         struct btrfs_trans_handle *trans;
1199         struct btrfs_extent_item *ei;
1200         struct btrfs_extent_inline_ref *iref;
1201         struct btrfs_extent_data_ref *dref;
1202         u64 bytenr;
1203         u64 num_bytes;
1204         u64 objectid;
1205         u64 last_byte;
1206         u64 first_free;
1207         u64 total_bytes;
1208         u32 sectorsize = root->sectorsize;
1209
1210         total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
1211         first_free =  BTRFS_SUPER_INFO_OFFSET + sectorsize * 2 - 1;
1212         first_free &= ~((u64)sectorsize - 1);
1213
1214         memset(&btrfs_inode, 0, sizeof(btrfs_inode));
1215         btrfs_set_stack_inode_generation(&btrfs_inode, 1);
1216         btrfs_set_stack_inode_size(&btrfs_inode, total_bytes);
1217         btrfs_set_stack_inode_nlink(&btrfs_inode, 1);
1218         btrfs_set_stack_inode_nbytes(&btrfs_inode, 0);
1219         btrfs_set_stack_inode_mode(&btrfs_inode, S_IFREG | 0400);
1220         btrfs_set_stack_inode_flags(&btrfs_inode, BTRFS_INODE_NODATASUM |
1221                                     BTRFS_INODE_READONLY);
1222         btrfs_init_path(&path);
1223         trans = btrfs_start_transaction(root, 1);
1224         BUG_ON(!trans);
1225
1226         objectid = btrfs_root_dirid(&root->root_item);
1227         ret = btrfs_find_free_objectid(trans, root, objectid, &objectid);
1228         if (ret)
1229                 goto fail;
1230
1231         /*
1232          * copy blocks covered by extent #0 to new positions. extent #0 is
1233          * special, we can't rely on relocate_extents_range to relocate it.
1234          */
1235         for (last_byte = 0; last_byte < first_free; last_byte += sectorsize) {
1236                 ret = custom_alloc_extent(root, sectorsize, 0, &key);
1237                 if (ret)
1238                         goto fail;
1239                 ret = copy_disk_extent(root, key.objectid, last_byte,
1240                                        sectorsize);
1241                 if (ret)
1242                         goto fail;
1243                 ret = btrfs_record_file_extent(trans, root, objectid,
1244                                                &btrfs_inode, last_byte,
1245                                                key.objectid, sectorsize);
1246                 if (ret)
1247                         goto fail;
1248         }
1249
1250         while(1) {
1251                 key.objectid = last_byte;
1252                 key.offset = 0;
1253                 btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
1254                 ret = btrfs_search_slot(trans, fs_info->extent_root,
1255                                         &key, &path, 0, 0);
1256                 if (ret < 0)
1257                         goto fail;
1258 next:
1259                 leaf = path.nodes[0];
1260                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1261                         ret = btrfs_next_leaf(extent_root, &path);
1262                         if (ret < 0)
1263                                 goto fail;
1264                         if (ret > 0)
1265                                 break;
1266                         leaf = path.nodes[0];
1267                 }
1268                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1269                 if (last_byte > key.objectid ||
1270                     key.type != BTRFS_EXTENT_ITEM_KEY) {
1271                         path.slots[0]++;
1272                         goto next;
1273                 }
1274
1275                 bytenr = key.objectid;
1276                 num_bytes = key.offset;
1277                 ei = btrfs_item_ptr(leaf, path.slots[0],
1278                                     struct btrfs_extent_item);
1279                 if (!(btrfs_extent_flags(leaf, ei) & BTRFS_EXTENT_FLAG_DATA)) {
1280                         path.slots[0]++;
1281                         goto next;
1282                 }
1283
1284                 BUG_ON(btrfs_item_size_nr(leaf, path.slots[0]) != sizeof(*ei) +
1285                        btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY));
1286
1287                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
1288                 key.type = btrfs_extent_inline_ref_type(leaf, iref);
1289                 BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
1290                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1291                 if (btrfs_extent_data_ref_root(leaf, dref) !=
1292                     BTRFS_FS_TREE_OBJECTID) {
1293                         path.slots[0]++;
1294                         goto next;
1295                 }
1296
1297                 if (bytenr > last_byte) {
1298                         ret = create_image_file_range(trans, root, objectid,
1299                                                       &btrfs_inode, last_byte,
1300                                                       bytenr, ext2_fs);
1301                         if (ret)
1302                                 goto fail;
1303                 }
1304                 ret = btrfs_record_file_extent(trans, root, objectid,
1305                                                &btrfs_inode, bytenr, bytenr,
1306                                                num_bytes);
1307                 if (ret)
1308                         goto fail;
1309                 last_byte = bytenr + num_bytes;
1310                 btrfs_release_path(&path);
1311
1312                 if (trans->blocks_used >= 4096) {
1313                         ret = btrfs_commit_transaction(trans, root);
1314                         BUG_ON(ret);
1315                         trans = btrfs_start_transaction(root, 1);
1316                         BUG_ON(!trans);
1317                 }
1318         }
1319         btrfs_release_path(&path);
1320         if (total_bytes > last_byte) {
1321                 ret = create_image_file_range(trans, root, objectid,
1322                                               &btrfs_inode, last_byte,
1323                                               total_bytes, ext2_fs);
1324                 if (ret)
1325                         goto fail;
1326         }
1327
1328         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1329         if (ret)
1330                 goto fail;
1331
1332         location.objectid = objectid;
1333         location.offset = 0;
1334         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1335         ret = btrfs_insert_dir_item(trans, root, name, strlen(name),
1336                                     btrfs_root_dirid(&root->root_item),
1337                                     &location, EXT2_FT_REG_FILE, objectid);
1338         if (ret)
1339                 goto fail;
1340         ret = btrfs_insert_inode_ref(trans, root, name, strlen(name),
1341                                      objectid,
1342                                      btrfs_root_dirid(&root->root_item),
1343                                      objectid);
1344         if (ret)
1345                 goto fail;
1346         location.objectid = btrfs_root_dirid(&root->root_item);
1347         location.offset = 0;
1348         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1349         ret = btrfs_lookup_inode(trans, root, &path, &location, 1);
1350         if (ret)
1351                 goto fail;
1352         leaf = path.nodes[0];
1353         inode_item = btrfs_item_ptr(leaf, path.slots[0],
1354                                     struct btrfs_inode_item);
1355         btrfs_set_inode_size(leaf, inode_item, strlen(name) * 2 +
1356                              btrfs_inode_size(leaf, inode_item));
1357         btrfs_mark_buffer_dirty(leaf);
1358         btrfs_release_path(&path);
1359         ret = btrfs_commit_transaction(trans, root);
1360         BUG_ON(ret);
1361 fail:
1362         btrfs_release_path(&path);
1363         return ret;
1364 }
1365
1366 static struct btrfs_root * link_subvol(struct btrfs_root *root,
1367                 const char *base, u64 root_objectid)
1368 {
1369         struct btrfs_trans_handle *trans;
1370         struct btrfs_fs_info *fs_info = root->fs_info;
1371         struct btrfs_root *tree_root = fs_info->tree_root;
1372         struct btrfs_root *new_root = NULL;
1373         struct btrfs_path *path;
1374         struct btrfs_inode_item *inode_item;
1375         struct extent_buffer *leaf;
1376         struct btrfs_key key;
1377         u64 dirid = btrfs_root_dirid(&root->root_item);
1378         u64 index = 2;
1379         char buf[BTRFS_NAME_LEN + 1]; /* for snprintf null */
1380         int len;
1381         int i;
1382         int ret;
1383
1384         len = strlen(base);
1385         if (len < 1 || len > BTRFS_NAME_LEN)
1386                 return NULL;
1387
1388         path = btrfs_alloc_path();
1389         BUG_ON(!path);
1390
1391         key.objectid = dirid;
1392         key.type = BTRFS_DIR_INDEX_KEY;
1393         key.offset = (u64)-1;
1394
1395         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1396         BUG_ON(ret <= 0);
1397
1398         if (path->slots[0] > 0) {
1399                 path->slots[0]--;
1400                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1401                 if (key.objectid == dirid && key.type == BTRFS_DIR_INDEX_KEY)
1402                         index = key.offset + 1;
1403         }
1404         btrfs_release_path(path);
1405
1406         trans = btrfs_start_transaction(root, 1);
1407         BUG_ON(!trans);
1408
1409         key.objectid = dirid;
1410         key.offset = 0;
1411         key.type =  BTRFS_INODE_ITEM_KEY;
1412
1413         ret = btrfs_lookup_inode(trans, root, path, &key, 1);
1414         BUG_ON(ret);
1415         leaf = path->nodes[0];
1416         inode_item = btrfs_item_ptr(leaf, path->slots[0],
1417                                     struct btrfs_inode_item);
1418
1419         key.objectid = root_objectid;
1420         key.offset = (u64)-1;
1421         key.type = BTRFS_ROOT_ITEM_KEY;
1422
1423         memcpy(buf, base, len);
1424         for (i = 0; i < 1024; i++) {
1425                 ret = btrfs_insert_dir_item(trans, root, buf, len,
1426                                             dirid, &key, BTRFS_FT_DIR, index);
1427                 if (ret != -EEXIST)
1428                         break;
1429                 len = snprintf(buf, ARRAY_SIZE(buf), "%s%d", base, i);
1430                 if (len < 1 || len > BTRFS_NAME_LEN) {
1431                         ret = -EINVAL;
1432                         break;
1433                 }
1434         }
1435         if (ret)
1436                 goto fail;
1437
1438         btrfs_set_inode_size(leaf, inode_item, len * 2 +
1439                              btrfs_inode_size(leaf, inode_item));
1440         btrfs_mark_buffer_dirty(leaf);
1441         btrfs_release_path(path);
1442
1443         /* add the backref first */
1444         ret = btrfs_add_root_ref(trans, tree_root, root_objectid,
1445                                  BTRFS_ROOT_BACKREF_KEY,
1446                                  root->root_key.objectid,
1447                                  dirid, index, buf, len);
1448         BUG_ON(ret);
1449
1450         /* now add the forward ref */
1451         ret = btrfs_add_root_ref(trans, tree_root, root->root_key.objectid,
1452                                  BTRFS_ROOT_REF_KEY, root_objectid,
1453                                  dirid, index, buf, len);
1454
1455         ret = btrfs_commit_transaction(trans, root);
1456         BUG_ON(ret);
1457
1458         new_root = btrfs_read_fs_root(fs_info, &key);
1459         if (IS_ERR(new_root))
1460                 new_root = NULL;
1461 fail:
1462         btrfs_free_path(path);
1463         return new_root;
1464 }
1465
1466 static int create_chunk_mapping(struct btrfs_trans_handle *trans,
1467                                 struct btrfs_root *root)
1468 {
1469         struct btrfs_fs_info *info = root->fs_info;
1470         struct btrfs_root *chunk_root = info->chunk_root;
1471         struct btrfs_root *extent_root = info->extent_root;
1472         struct btrfs_device *device;
1473         struct btrfs_block_group_cache *cache;
1474         struct btrfs_dev_extent *extent;
1475         struct extent_buffer *leaf;
1476         struct btrfs_chunk chunk;
1477         struct btrfs_key key;
1478         struct btrfs_path path;
1479         u64 cur_start;
1480         u64 total_bytes;
1481         u64 chunk_objectid;
1482         int ret;
1483
1484         btrfs_init_path(&path);
1485
1486         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
1487         chunk_objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
1488
1489         BUG_ON(list_empty(&info->fs_devices->devices));
1490         device = list_entry(info->fs_devices->devices.next,
1491                             struct btrfs_device, dev_list);
1492         BUG_ON(device->devid != info->fs_devices->latest_devid);
1493
1494         /* delete device extent created by make_btrfs */
1495         key.objectid = device->devid;
1496         key.offset = 0;
1497         key.type = BTRFS_DEV_EXTENT_KEY;
1498         ret = btrfs_search_slot(trans, device->dev_root, &key, &path, -1, 1);
1499         if (ret < 0)
1500                 goto err;
1501
1502         BUG_ON(ret > 0);
1503         ret = btrfs_del_item(trans, device->dev_root, &path);
1504         if (ret)
1505                 goto err;
1506         btrfs_release_path(&path);
1507
1508         /* delete chunk item created by make_btrfs */
1509         key.objectid = chunk_objectid;
1510         key.offset = 0;
1511         key.type = BTRFS_CHUNK_ITEM_KEY;
1512         ret = btrfs_search_slot(trans, chunk_root, &key, &path, -1, 1);
1513         if (ret < 0)
1514                 goto err;
1515
1516         BUG_ON(ret > 0);
1517         ret = btrfs_del_item(trans, chunk_root, &path);
1518         if (ret)
1519                 goto err;
1520         btrfs_release_path(&path);
1521
1522         /* for each block group, create device extent and chunk item */
1523         cur_start = 0;
1524         while (cur_start < total_bytes) {
1525                 cache = btrfs_lookup_block_group(root->fs_info, cur_start);
1526                 BUG_ON(!cache);
1527
1528                 /* insert device extent */
1529                 key.objectid = device->devid;
1530                 key.offset = cache->key.objectid;
1531                 key.type = BTRFS_DEV_EXTENT_KEY;
1532                 ret = btrfs_insert_empty_item(trans, device->dev_root, &path,
1533                                               &key, sizeof(*extent));
1534                 if (ret)
1535                         goto err;
1536
1537                 leaf = path.nodes[0];
1538                 extent = btrfs_item_ptr(leaf, path.slots[0],
1539                                         struct btrfs_dev_extent);
1540
1541                 btrfs_set_dev_extent_chunk_tree(leaf, extent,
1542                                                 chunk_root->root_key.objectid);
1543                 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
1544                                                     chunk_objectid);
1545                 btrfs_set_dev_extent_chunk_offset(leaf, extent,
1546                                                   cache->key.objectid);
1547                 btrfs_set_dev_extent_length(leaf, extent, cache->key.offset);
1548                 write_extent_buffer(leaf, root->fs_info->chunk_tree_uuid,
1549                     (unsigned long)btrfs_dev_extent_chunk_tree_uuid(extent),
1550                     BTRFS_UUID_SIZE);
1551                 btrfs_mark_buffer_dirty(leaf);
1552                 btrfs_release_path(&path);
1553
1554                 /* insert chunk item */
1555                 btrfs_set_stack_chunk_length(&chunk, cache->key.offset);
1556                 btrfs_set_stack_chunk_owner(&chunk,
1557                                             extent_root->root_key.objectid);
1558                 btrfs_set_stack_chunk_stripe_len(&chunk, BTRFS_STRIPE_LEN);
1559                 btrfs_set_stack_chunk_type(&chunk, cache->flags);
1560                 btrfs_set_stack_chunk_io_align(&chunk, device->io_align);
1561                 btrfs_set_stack_chunk_io_width(&chunk, device->io_width);
1562                 btrfs_set_stack_chunk_sector_size(&chunk, device->sector_size);
1563                 btrfs_set_stack_chunk_num_stripes(&chunk, 1);
1564                 btrfs_set_stack_chunk_sub_stripes(&chunk, 0);
1565                 btrfs_set_stack_stripe_devid(&chunk.stripe, device->devid);
1566                 btrfs_set_stack_stripe_offset(&chunk.stripe,
1567                                               cache->key.objectid);
1568                 memcpy(&chunk.stripe.dev_uuid, device->uuid, BTRFS_UUID_SIZE);
1569
1570                 key.objectid = chunk_objectid;
1571                 key.offset = cache->key.objectid;
1572                 key.type = BTRFS_CHUNK_ITEM_KEY;
1573
1574                 ret = btrfs_insert_item(trans, chunk_root, &key, &chunk,
1575                                         btrfs_chunk_item_size(1));
1576                 if (ret)
1577                         goto err;
1578
1579                 cur_start = cache->key.objectid + cache->key.offset;
1580         }
1581
1582         device->bytes_used = total_bytes;
1583         ret = btrfs_update_device(trans, device);
1584 err:
1585         btrfs_release_path(&path);
1586         return ret;
1587 }
1588
1589 static int create_subvol(struct btrfs_trans_handle *trans,
1590                          struct btrfs_root *root, u64 root_objectid)
1591 {
1592         struct extent_buffer *tmp;
1593         struct btrfs_root *new_root;
1594         struct btrfs_key key;
1595         struct btrfs_root_item root_item;
1596         int ret;
1597
1598         ret = btrfs_copy_root(trans, root, root->node, &tmp,
1599                               root_objectid);
1600         BUG_ON(ret);
1601
1602         memcpy(&root_item, &root->root_item, sizeof(root_item));
1603         btrfs_set_root_bytenr(&root_item, tmp->start);
1604         btrfs_set_root_level(&root_item, btrfs_header_level(tmp));
1605         btrfs_set_root_generation(&root_item, trans->transid);
1606         free_extent_buffer(tmp);
1607
1608         key.objectid = root_objectid;
1609         key.type = BTRFS_ROOT_ITEM_KEY;
1610         key.offset = trans->transid;
1611         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1612                                 &key, &root_item);
1613
1614         key.offset = (u64)-1;
1615         new_root = btrfs_read_fs_root(root->fs_info, &key);
1616         BUG_ON(!new_root || IS_ERR(new_root));
1617
1618         ret = btrfs_make_root_dir(trans, new_root, BTRFS_FIRST_FREE_OBJECTID);
1619         BUG_ON(ret);
1620
1621         return 0;
1622 }
1623
1624 static int init_btrfs(struct btrfs_root *root)
1625 {
1626         int ret;
1627         struct btrfs_key location;
1628         struct btrfs_trans_handle *trans;
1629         struct btrfs_fs_info *fs_info = root->fs_info;
1630         struct extent_buffer *tmp;
1631
1632         trans = btrfs_start_transaction(root, 1);
1633         BUG_ON(!trans);
1634         ret = btrfs_make_block_groups(trans, root);
1635         if (ret)
1636                 goto err;
1637         ret = btrfs_fix_block_accounting(trans, root);
1638         if (ret)
1639                 goto err;
1640         ret = create_chunk_mapping(trans, root);
1641         if (ret)
1642                 goto err;
1643         ret = btrfs_make_root_dir(trans, fs_info->tree_root,
1644                                   BTRFS_ROOT_TREE_DIR_OBJECTID);
1645         if (ret)
1646                 goto err;
1647         memcpy(&location, &root->root_key, sizeof(location));
1648         location.offset = (u64)-1;
1649         ret = btrfs_insert_dir_item(trans, fs_info->tree_root, "default", 7,
1650                                 btrfs_super_root_dir(fs_info->super_copy),
1651                                 &location, BTRFS_FT_DIR, 0);
1652         if (ret)
1653                 goto err;
1654         ret = btrfs_insert_inode_ref(trans, fs_info->tree_root, "default", 7,
1655                                 location.objectid,
1656                                 btrfs_super_root_dir(fs_info->super_copy), 0);
1657         if (ret)
1658                 goto err;
1659         btrfs_set_root_dirid(&fs_info->fs_root->root_item,
1660                              BTRFS_FIRST_FREE_OBJECTID);
1661
1662         /* subvol for ext2 image file */
1663         ret = create_subvol(trans, root, EXT2_IMAGE_SUBVOL_OBJECTID);
1664         BUG_ON(ret);
1665         /* subvol for data relocation */
1666         ret = create_subvol(trans, root, BTRFS_DATA_RELOC_TREE_OBJECTID);
1667         BUG_ON(ret);
1668
1669         extent_buffer_get(fs_info->csum_root->node);
1670         ret = __btrfs_cow_block(trans, fs_info->csum_root,
1671                                 fs_info->csum_root->node, NULL, 0, &tmp, 0, 0);
1672         BUG_ON(ret);
1673         free_extent_buffer(tmp);
1674
1675         ret = btrfs_commit_transaction(trans, root);
1676         BUG_ON(ret);
1677 err:
1678         return ret;
1679 }
1680
1681 /*
1682  * Migrate super block to its default position and zero 0 ~ 16k
1683  */
1684 static int migrate_super_block(int fd, u64 old_bytenr, u32 sectorsize)
1685 {
1686         int ret;
1687         struct extent_buffer *buf;
1688         struct btrfs_super_block *super;
1689         u32 len;
1690         u32 bytenr;
1691
1692         BUG_ON(sectorsize < sizeof(*super));
1693         buf = malloc(sizeof(*buf) + sectorsize);
1694         if (!buf)
1695                 return -ENOMEM;
1696
1697         buf->len = sectorsize;
1698         ret = pread(fd, buf->data, sectorsize, old_bytenr);
1699         if (ret != sectorsize)
1700                 goto fail;
1701
1702         super = (struct btrfs_super_block *)buf->data;
1703         BUG_ON(btrfs_super_bytenr(super) != old_bytenr);
1704         btrfs_set_super_bytenr(super, BTRFS_SUPER_INFO_OFFSET);
1705
1706         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1707         ret = pwrite(fd, buf->data, sectorsize, BTRFS_SUPER_INFO_OFFSET);
1708         if (ret != sectorsize)
1709                 goto fail;
1710
1711         ret = fsync(fd);
1712         if (ret)
1713                 goto fail;
1714
1715         memset(buf->data, 0, sectorsize);
1716         for (bytenr = 0; bytenr < BTRFS_SUPER_INFO_OFFSET; ) {
1717                 len = BTRFS_SUPER_INFO_OFFSET - bytenr;
1718                 if (len > sectorsize)
1719                         len = sectorsize;
1720                 ret = pwrite(fd, buf->data, len, bytenr);
1721                 if (ret != len) {
1722                         fprintf(stderr, "unable to zero fill device\n");
1723                         break;
1724                 }
1725                 bytenr += len;
1726         }
1727         ret = 0;
1728         fsync(fd);
1729 fail:
1730         free(buf);
1731         if (ret > 0)
1732                 ret = -1;
1733         return ret;
1734 }
1735
1736 static int prepare_system_chunk_sb(struct btrfs_super_block *super)
1737 {
1738         struct btrfs_chunk *chunk;
1739         struct btrfs_disk_key *key;
1740         u32 sectorsize = btrfs_super_sectorsize(super);
1741
1742         key = (struct btrfs_disk_key *)(super->sys_chunk_array);
1743         chunk = (struct btrfs_chunk *)(super->sys_chunk_array +
1744                                        sizeof(struct btrfs_disk_key));
1745
1746         btrfs_set_disk_key_objectid(key, BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1747         btrfs_set_disk_key_type(key, BTRFS_CHUNK_ITEM_KEY);
1748         btrfs_set_disk_key_offset(key, 0);
1749
1750         btrfs_set_stack_chunk_length(chunk, btrfs_super_total_bytes(super));
1751         btrfs_set_stack_chunk_owner(chunk, BTRFS_EXTENT_TREE_OBJECTID);
1752         btrfs_set_stack_chunk_stripe_len(chunk, BTRFS_STRIPE_LEN);
1753         btrfs_set_stack_chunk_type(chunk, BTRFS_BLOCK_GROUP_SYSTEM);
1754         btrfs_set_stack_chunk_io_align(chunk, sectorsize);
1755         btrfs_set_stack_chunk_io_width(chunk, sectorsize);
1756         btrfs_set_stack_chunk_sector_size(chunk, sectorsize);
1757         btrfs_set_stack_chunk_num_stripes(chunk, 1);
1758         btrfs_set_stack_chunk_sub_stripes(chunk, 0);
1759         chunk->stripe.devid = super->dev_item.devid;
1760         btrfs_set_stack_stripe_offset(&chunk->stripe, 0);
1761         memcpy(chunk->stripe.dev_uuid, super->dev_item.uuid, BTRFS_UUID_SIZE);
1762         btrfs_set_super_sys_array_size(super, sizeof(*key) + sizeof(*chunk));
1763         return 0;
1764 }
1765
1766 static int prepare_system_chunk(int fd, u64 sb_bytenr, u32 sectorsize)
1767 {
1768         int ret;
1769         struct extent_buffer *buf;
1770         struct btrfs_super_block *super;
1771
1772         BUG_ON(sectorsize < sizeof(*super));
1773         buf = malloc(sizeof(*buf) + sectorsize);
1774         if (!buf)
1775                 return -ENOMEM;
1776
1777         buf->len = sectorsize;
1778         ret = pread(fd, buf->data, sectorsize, sb_bytenr);
1779         if (ret != sectorsize)
1780                 goto fail;
1781
1782         super = (struct btrfs_super_block *)buf->data;
1783         BUG_ON(btrfs_super_bytenr(super) != sb_bytenr);
1784         BUG_ON(btrfs_super_num_devices(super) != 1);
1785
1786         ret = prepare_system_chunk_sb(super);
1787         if (ret)
1788                 goto fail;
1789
1790         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1791         ret = pwrite(fd, buf->data, sectorsize, sb_bytenr);
1792         if (ret != sectorsize)
1793                 goto fail;
1794
1795         ret = 0;
1796 fail:
1797         free(buf);
1798         if (ret > 0)
1799                 ret = -1;
1800         return ret;
1801 }
1802
1803 static int relocate_one_reference(struct btrfs_trans_handle *trans,
1804                                   struct btrfs_root *root,
1805                                   u64 extent_start, u64 extent_size,
1806                                   struct btrfs_key *extent_key,
1807                                   struct extent_io_tree *reloc_tree)
1808 {
1809         struct extent_buffer *leaf;
1810         struct btrfs_file_extent_item *fi;
1811         struct btrfs_key key;
1812         struct btrfs_path path;
1813         struct btrfs_inode_item inode;
1814         struct blk_iterate_data data;
1815         u64 bytenr;
1816         u64 num_bytes;
1817         u64 cur_offset;
1818         u64 new_pos;
1819         u64 nbytes;
1820         u64 sector_end;
1821         u32 sectorsize = root->sectorsize;
1822         unsigned long ptr;
1823         int datacsum;
1824         int fd;
1825         int ret;
1826
1827         btrfs_init_path(&path);
1828         ret = btrfs_search_slot(trans, root, extent_key, &path, -1, 1);
1829         if (ret)
1830                 goto fail;
1831
1832         leaf = path.nodes[0];
1833         fi = btrfs_item_ptr(leaf, path.slots[0],
1834                             struct btrfs_file_extent_item);
1835         BUG_ON(btrfs_file_extent_offset(leaf, fi) > 0);
1836         if (extent_start != btrfs_file_extent_disk_bytenr(leaf, fi) ||
1837             extent_size != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1838                 ret = 1;
1839                 goto fail;
1840         }
1841
1842         bytenr = extent_start + btrfs_file_extent_offset(leaf, fi);
1843         num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
1844
1845         ret = btrfs_del_item(trans, root, &path);
1846         if (ret)
1847                 goto fail;
1848
1849         ret = btrfs_free_extent(trans, root, extent_start, extent_size, 0,
1850                                 root->root_key.objectid,
1851                                 extent_key->objectid, extent_key->offset);
1852         if (ret)
1853                 goto fail;
1854
1855         btrfs_release_path(&path);
1856
1857         key.objectid = extent_key->objectid;
1858         key.offset = 0;
1859         key.type =  BTRFS_INODE_ITEM_KEY;
1860         ret = btrfs_lookup_inode(trans, root, &path, &key, 0);
1861         if (ret)
1862                 goto fail;
1863
1864         leaf = path.nodes[0];
1865         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1866         read_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1867         btrfs_release_path(&path);
1868
1869         BUG_ON(num_bytes & (sectorsize - 1));
1870         nbytes = btrfs_stack_inode_nbytes(&inode) - num_bytes;
1871         btrfs_set_stack_inode_nbytes(&inode, nbytes);
1872         datacsum = !(btrfs_stack_inode_flags(&inode) & BTRFS_INODE_NODATASUM);
1873
1874         data = (struct blk_iterate_data) {
1875                 .trans          = trans,
1876                 .root           = root,
1877                 .inode          = &inode,
1878                 .objectid       = extent_key->objectid,
1879                 .first_block    = extent_key->offset / sectorsize,
1880                 .disk_block     = 0,
1881                 .num_blocks     = 0,
1882                 .boundary       = (u64)-1,
1883                 .checksum       = datacsum,
1884                 .errcode        = 0,
1885         };
1886
1887         cur_offset = extent_key->offset;
1888         while (num_bytes > 0) {
1889                 sector_end = bytenr + sectorsize - 1;
1890                 if (test_range_bit(reloc_tree, bytenr, sector_end,
1891                                    EXTENT_LOCKED, 1)) {
1892                         ret = get_state_private(reloc_tree, bytenr, &new_pos);
1893                         BUG_ON(ret);
1894                 } else {
1895                         ret = custom_alloc_extent(root, sectorsize, 0, &key);
1896                         if (ret)
1897                                 goto fail;
1898                         new_pos = key.objectid;
1899
1900                         if (cur_offset == extent_key->offset) {
1901                                 fd = root->fs_info->fs_devices->latest_bdev;
1902                                 readahead(fd, bytenr, num_bytes);
1903                         }
1904                         ret = copy_disk_extent(root, new_pos, bytenr,
1905                                                sectorsize);
1906                         if (ret)
1907                                 goto fail;
1908                         ret = set_extent_bits(reloc_tree, bytenr, sector_end,
1909                                               EXTENT_LOCKED, GFP_NOFS);
1910                         BUG_ON(ret);
1911                         ret = set_state_private(reloc_tree, bytenr, new_pos);
1912                         BUG_ON(ret);
1913                 }
1914
1915                 ret = block_iterate_proc(NULL, new_pos / sectorsize,
1916                                          cur_offset / sectorsize, &data);
1917                 if (ret & BLOCK_ABORT) {
1918                         ret = data.errcode;
1919                         goto fail;
1920                 }
1921
1922                 cur_offset += sectorsize;
1923                 bytenr += sectorsize;
1924                 num_bytes -= sectorsize;
1925         }
1926
1927         if (data.num_blocks > 0) {
1928                 ret = record_file_blocks(trans, root,
1929                                          extent_key->objectid, &inode,
1930                                          data.first_block, data.disk_block,
1931                                          data.num_blocks, datacsum);
1932                 if (ret)
1933                         goto fail;
1934         }
1935
1936         key.objectid = extent_key->objectid;
1937         key.offset = 0;
1938         key.type =  BTRFS_INODE_ITEM_KEY;
1939         ret = btrfs_lookup_inode(trans, root, &path, &key, 1);
1940         if (ret)
1941                 goto fail;
1942
1943         leaf = path.nodes[0];
1944         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1945         write_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1946         btrfs_mark_buffer_dirty(leaf);
1947         btrfs_release_path(&path);
1948
1949 fail:
1950         btrfs_release_path(&path);
1951         return ret;
1952 }
1953
1954 static int relocate_extents_range(struct btrfs_root *fs_root,
1955                                   struct btrfs_root *ext2_root,
1956                                   u64 start_byte, u64 end_byte)
1957 {
1958         struct btrfs_fs_info *info = fs_root->fs_info;
1959         struct btrfs_root *extent_root = info->extent_root;
1960         struct btrfs_root *cur_root = NULL;
1961         struct btrfs_trans_handle *trans;
1962         struct btrfs_extent_data_ref *dref;
1963         struct btrfs_extent_inline_ref *iref;
1964         struct btrfs_extent_item *ei;
1965         struct extent_buffer *leaf;
1966         struct btrfs_key key;
1967         struct btrfs_key extent_key;
1968         struct btrfs_path path;
1969         struct extent_io_tree reloc_tree;
1970         unsigned long ptr;
1971         unsigned long end;
1972         u64 cur_byte;
1973         u64 num_bytes;
1974         u64 ref_root;
1975         u64 num_extents;
1976         int pass = 0;
1977         int ret;
1978
1979         btrfs_init_path(&path);
1980         extent_io_tree_init(&reloc_tree);
1981
1982         key.objectid = start_byte;
1983         key.offset = 0;
1984         key.type = BTRFS_EXTENT_ITEM_KEY;
1985         ret = btrfs_search_slot(NULL, extent_root, &key, &path, 0, 0);
1986         if (ret < 0)
1987                 goto fail;
1988         if (ret > 0) {
1989                 ret = btrfs_previous_item(extent_root, &path, 0,
1990                                           BTRFS_EXTENT_ITEM_KEY);
1991                 if (ret < 0)
1992                         goto fail;
1993                 if (ret == 0) {
1994                         leaf = path.nodes[0];
1995                         btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1996                         if (key.objectid + key.offset > start_byte)
1997                                 start_byte = key.objectid;
1998                 }
1999         }
2000         btrfs_release_path(&path);
2001 again:
2002         cur_root = (pass % 2 == 0) ? ext2_root : fs_root;
2003         num_extents = 0;
2004
2005         trans = btrfs_start_transaction(cur_root, 1);
2006         BUG_ON(!trans);
2007
2008         cur_byte = start_byte;
2009         while (1) {
2010                 key.objectid = cur_byte;
2011                 key.offset = 0;
2012                 key.type = BTRFS_EXTENT_ITEM_KEY;
2013                 ret = btrfs_search_slot(trans, extent_root,
2014                                         &key, &path, 0, 0);
2015                 if (ret < 0)
2016                         goto fail;
2017 next:
2018                 leaf = path.nodes[0];
2019                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2020                         ret = btrfs_next_leaf(extent_root, &path);
2021                         if (ret < 0)
2022                                 goto fail;
2023                         if (ret > 0)
2024                                 break;
2025                         leaf = path.nodes[0];
2026                 }
2027
2028                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2029                 if (key.objectid < cur_byte ||
2030                     key.type != BTRFS_EXTENT_ITEM_KEY) {
2031                         path.slots[0]++;
2032                         goto next;
2033                 }
2034                 if (key.objectid >= end_byte)
2035                         break;
2036
2037                 num_extents++;
2038
2039                 cur_byte = key.objectid;
2040                 num_bytes = key.offset;
2041                 ei = btrfs_item_ptr(leaf, path.slots[0],
2042                                     struct btrfs_extent_item);
2043                 BUG_ON(!(btrfs_extent_flags(leaf, ei) &
2044                          BTRFS_EXTENT_FLAG_DATA));
2045
2046                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2047                 end = ptr + btrfs_item_size_nr(leaf, path.slots[0]);
2048
2049                 ptr += sizeof(struct btrfs_extent_item);
2050
2051                 while (ptr < end) {
2052                         iref = (struct btrfs_extent_inline_ref *)ptr;
2053                         key.type = btrfs_extent_inline_ref_type(leaf, iref);
2054                         BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
2055                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2056                         ref_root = btrfs_extent_data_ref_root(leaf, dref);
2057                         extent_key.objectid =
2058                                 btrfs_extent_data_ref_objectid(leaf, dref);
2059                         extent_key.offset =
2060                                 btrfs_extent_data_ref_offset(leaf, dref);
2061                         extent_key.type = BTRFS_EXTENT_DATA_KEY;
2062                         BUG_ON(btrfs_extent_data_ref_count(leaf, dref) != 1);
2063
2064                         if (ref_root == cur_root->root_key.objectid)
2065                                 break;
2066
2067                         ptr += btrfs_extent_inline_ref_size(key.type);
2068                 }
2069
2070                 if (ptr >= end) {
2071                         path.slots[0]++;
2072                         goto next;
2073                 }
2074
2075                 ret = relocate_one_reference(trans, cur_root, cur_byte,
2076                                              num_bytes, &extent_key,
2077                                              &reloc_tree);
2078                 if (ret < 0)
2079                         goto fail;
2080
2081                 cur_byte += num_bytes;
2082                 btrfs_release_path(&path);
2083
2084                 if (trans->blocks_used >= 4096) {
2085                         ret = btrfs_commit_transaction(trans, cur_root);
2086                         BUG_ON(ret);
2087                         trans = btrfs_start_transaction(cur_root, 1);
2088                         BUG_ON(!trans);
2089                 }
2090         }
2091         btrfs_release_path(&path);
2092
2093         ret = btrfs_commit_transaction(trans, cur_root);
2094         BUG_ON(ret);
2095
2096         if (num_extents > 0 && pass++ < 16)
2097                 goto again;
2098
2099         ret = (num_extents > 0) ? -1 : 0;
2100 fail:
2101         btrfs_release_path(&path);
2102         extent_io_tree_cleanup(&reloc_tree);
2103         return ret;
2104 }
2105
2106 /*
2107  * relocate data in system chunk
2108  */
2109 static int cleanup_sys_chunk(struct btrfs_root *fs_root,
2110                              struct btrfs_root *ext2_root)
2111 {
2112         struct btrfs_block_group_cache *cache;
2113         int i, ret = 0;
2114         u64 offset = 0;
2115         u64 end_byte;
2116
2117         while(1) {
2118                 cache = btrfs_lookup_block_group(fs_root->fs_info, offset);
2119                 if (!cache)
2120                         break;
2121
2122                 end_byte = cache->key.objectid + cache->key.offset;
2123                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2124                         ret = relocate_extents_range(fs_root, ext2_root,
2125                                                      cache->key.objectid,
2126                                                      end_byte);
2127                         if (ret)
2128                                 goto fail;
2129                 }
2130                 offset = end_byte;
2131         }
2132         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2133                 offset = btrfs_sb_offset(i);
2134                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
2135
2136                 ret = relocate_extents_range(fs_root, ext2_root,
2137                                              offset, offset + BTRFS_STRIPE_LEN);
2138                 if (ret)
2139                         goto fail;
2140         }
2141         ret = 0;
2142 fail:
2143         return ret;
2144 }
2145
2146 static int fixup_chunk_mapping(struct btrfs_root *root)
2147 {
2148         struct btrfs_trans_handle *trans;
2149         struct btrfs_fs_info *info = root->fs_info;
2150         struct btrfs_root *chunk_root = info->chunk_root;
2151         struct extent_buffer *leaf;
2152         struct btrfs_key key;
2153         struct btrfs_path path;
2154         struct btrfs_chunk chunk;
2155         unsigned long ptr;
2156         u32 size;
2157         u64 type;
2158         int ret;
2159
2160         btrfs_init_path(&path);
2161
2162         trans = btrfs_start_transaction(root, 1);
2163         BUG_ON(!trans);
2164
2165         /*
2166          * recow the whole chunk tree. this will move all chunk tree blocks
2167          * into system block group.
2168          */
2169         memset(&key, 0, sizeof(key));
2170         while (1) {
2171                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2172                 if (ret < 0)
2173                         goto err;
2174
2175                 ret = btrfs_next_leaf(chunk_root, &path);
2176                 if (ret < 0)
2177                         goto err;
2178                 if (ret > 0)
2179                         break;
2180
2181                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2182                 btrfs_release_path(&path);
2183         }
2184         btrfs_release_path(&path);
2185
2186         /* fixup the system chunk array in super block */
2187         btrfs_set_super_sys_array_size(info->super_copy, 0);
2188
2189         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2190         key.offset = 0;
2191         key.type = BTRFS_CHUNK_ITEM_KEY;
2192
2193         ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 0);
2194         if (ret < 0)
2195                 goto err;
2196         BUG_ON(ret != 0);
2197         while(1) {
2198                 leaf = path.nodes[0];
2199                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2200                         ret = btrfs_next_leaf(chunk_root, &path);
2201                         if (ret < 0)
2202                                 goto err;
2203                         if (ret > 0)
2204                                 break;
2205                         leaf = path.nodes[0];
2206                 }
2207                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2208                 if (key.type != BTRFS_CHUNK_ITEM_KEY)
2209                         goto next;
2210
2211                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2212                 size = btrfs_item_size_nr(leaf, path.slots[0]);
2213                 BUG_ON(size != sizeof(chunk));
2214                 read_extent_buffer(leaf, &chunk, ptr, size);
2215                 type = btrfs_stack_chunk_type(&chunk);
2216
2217                 if (!(type & BTRFS_BLOCK_GROUP_SYSTEM))
2218                         goto next;
2219
2220                 ret = btrfs_add_system_chunk(trans, chunk_root, &key,
2221                                              &chunk, size);
2222                 if (ret)
2223                         goto err;
2224 next:
2225                 path.slots[0]++;
2226         }
2227
2228         ret = btrfs_commit_transaction(trans, root);
2229         BUG_ON(ret);
2230 err:
2231         btrfs_release_path(&path);
2232         return ret;
2233 }
2234
2235 static int do_convert(const char *devname, int datacsum, int packing, int noxattr,
2236                int copylabel, const char *fslabel, int progress)
2237 {
2238         int i, ret;
2239         int fd = -1;
2240         u32 blocksize;
2241         u64 blocks[7];
2242         u64 total_bytes;
2243         u64 super_bytenr;
2244         ext2_filsys ext2_fs;
2245         struct btrfs_root *root;
2246         struct btrfs_root *ext2_root;
2247         struct task_ctx ctx;
2248
2249         ret = open_ext2fs(devname, &ext2_fs);
2250         if (ret) {
2251                 fprintf(stderr, "unable to open the Ext2fs\n");
2252                 goto fail;
2253         }
2254         blocksize = ext2_fs->blocksize;
2255         total_bytes = (u64)ext2_fs->super->s_blocks_count * blocksize;
2256         if (blocksize < 4096) {
2257                 fprintf(stderr, "block size is too small\n");
2258                 goto fail;
2259         }
2260         if (!(ext2_fs->super->s_feature_incompat &
2261               EXT2_FEATURE_INCOMPAT_FILETYPE)) {
2262                 fprintf(stderr, "filetype feature is missing\n");
2263                 goto fail;
2264         }
2265         for (i = 0; i < 7; i++) {
2266                 ret = ext2_alloc_block(ext2_fs, 0, blocks + i);
2267                 if (ret) {
2268                         fprintf(stderr, "not enough free space\n");
2269                         goto fail;
2270                 }
2271                 blocks[i] *= blocksize;
2272         }
2273         super_bytenr = blocks[0];
2274         fd = open(devname, O_RDWR);
2275         if (fd < 0) {
2276                 fprintf(stderr, "unable to open %s\n", devname);
2277                 goto fail;
2278         }
2279         ret = make_btrfs(fd, devname, ext2_fs->super->s_volume_name,
2280                          NULL, blocks, total_bytes, blocksize, blocksize,
2281                          blocksize, blocksize, 0);
2282         if (ret) {
2283                 fprintf(stderr, "unable to create initial ctree: %s\n",
2284                         strerror(-ret));
2285                 goto fail;
2286         }
2287         /* create a system chunk that maps the whole device */
2288         ret = prepare_system_chunk(fd, super_bytenr, blocksize);
2289         if (ret) {
2290                 fprintf(stderr, "unable to update system chunk\n");
2291                 goto fail;
2292         }
2293         root = open_ctree_fd(fd, devname, super_bytenr, OPEN_CTREE_WRITES);
2294         if (!root) {
2295                 fprintf(stderr, "unable to open ctree\n");
2296                 goto fail;
2297         }
2298         ret = cache_free_extents(root, ext2_fs);
2299         if (ret) {
2300                 fprintf(stderr, "error during cache_free_extents %d\n", ret);
2301                 goto fail;
2302         }
2303         root->fs_info->extent_ops = &extent_ops;
2304         /* recover block allocation bitmap */
2305         for (i = 0; i < 7; i++) {
2306                 blocks[i] /= blocksize;
2307                 ext2_free_block(ext2_fs, blocks[i]);
2308         }
2309         ret = init_btrfs(root);
2310         if (ret) {
2311                 fprintf(stderr, "unable to setup the root tree\n");
2312                 goto fail;
2313         }
2314         printf("creating btrfs metadata.\n");
2315         ctx.max_copy_inodes = (ext2_fs->super->s_inodes_count
2316                         - ext2_fs->super->s_free_inodes_count);
2317         ctx.cur_copy_inodes = 0;
2318
2319         if (progress) {
2320                 ctx.info = task_init(print_copied_inodes, after_copied_inodes, &ctx);
2321                 task_start(ctx.info);
2322         }
2323         ret = copy_inodes(root, ext2_fs, datacsum, packing, noxattr, &ctx);
2324         if (ret) {
2325                 fprintf(stderr, "error during copy_inodes %d\n", ret);
2326                 goto fail;
2327         }
2328         if (progress) {
2329                 task_stop(ctx.info);
2330                 task_deinit(ctx.info);
2331         }
2332         printf("creating ext2fs image file.\n");
2333         ext2_root = link_subvol(root, "ext2_saved", EXT2_IMAGE_SUBVOL_OBJECTID);
2334         if (!ext2_root) {
2335                 fprintf(stderr, "unable to create subvol\n");
2336                 goto fail;
2337         }
2338         ret = create_ext2_image(ext2_root, ext2_fs, "image");
2339         if (ret) {
2340                 fprintf(stderr, "error during create_ext2_image %d\n", ret);
2341                 goto fail;
2342         }
2343         memset(root->fs_info->super_copy->label, 0, BTRFS_LABEL_SIZE);
2344         if (copylabel == 1) {
2345                 strncpy(root->fs_info->super_copy->label,
2346                                 ext2_fs->super->s_volume_name, 16);
2347                 fprintf(stderr, "copy label '%s'\n",
2348                                 root->fs_info->super_copy->label);
2349         } else if (copylabel == -1) {
2350                 strncpy(root->fs_info->super_copy->label, fslabel, BTRFS_LABEL_SIZE);
2351                 fprintf(stderr, "set label to '%s'\n", fslabel);
2352         }
2353
2354         printf("cleaning up system chunk.\n");
2355         ret = cleanup_sys_chunk(root, ext2_root);
2356         if (ret) {
2357                 fprintf(stderr, "error during cleanup_sys_chunk %d\n", ret);
2358                 goto fail;
2359         }
2360         ret = close_ctree(root);
2361         if (ret) {
2362                 fprintf(stderr, "error during close_ctree %d\n", ret);
2363                 goto fail;
2364         }
2365         close_ext2fs(ext2_fs);
2366
2367         /*
2368          * If this step succeed, we get a mountable btrfs. Otherwise
2369          * the ext2fs is left unchanged.
2370          */
2371         ret = migrate_super_block(fd, super_bytenr, blocksize);
2372         if (ret) {
2373                 fprintf(stderr, "unable to migrate super block\n");
2374                 goto fail;
2375         }
2376
2377         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2378         if (!root) {
2379                 fprintf(stderr, "unable to open ctree\n");
2380                 goto fail;
2381         }
2382         /* move chunk tree into system chunk. */
2383         ret = fixup_chunk_mapping(root);
2384         if (ret) {
2385                 fprintf(stderr, "error during fixup_chunk_tree\n");
2386                 goto fail;
2387         }
2388         ret = close_ctree(root);
2389         close(fd);
2390
2391         printf("conversion complete.\n");
2392         return 0;
2393 fail:
2394         if (fd != -1)
2395                 close(fd);
2396         fprintf(stderr, "conversion aborted.\n");
2397         return -1;
2398 }
2399
2400 static int may_rollback(struct btrfs_root *root)
2401 {
2402         struct btrfs_fs_info *info = root->fs_info;
2403         struct btrfs_multi_bio *multi = NULL;
2404         u64 bytenr;
2405         u64 length;
2406         u64 physical;
2407         u64 total_bytes;
2408         int num_stripes;
2409         int ret;
2410
2411         if (btrfs_super_num_devices(info->super_copy) != 1)
2412                 goto fail;
2413
2414         bytenr = BTRFS_SUPER_INFO_OFFSET;
2415         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
2416
2417         while (1) {
2418                 ret = btrfs_map_block(&info->mapping_tree, WRITE, bytenr,
2419                                       &length, &multi, 0, NULL);
2420                 if (ret) {
2421                         if (ret == -ENOENT) {
2422                                 /* removed block group at the tail */
2423                                 if (length == (u64)-1)
2424                                         break;
2425
2426                                 /* removed block group in the middle */
2427                                 goto next;
2428                         }
2429                         goto fail;
2430                 }
2431
2432                 num_stripes = multi->num_stripes;
2433                 physical = multi->stripes[0].physical;
2434                 kfree(multi);
2435
2436                 if (num_stripes != 1 || physical != bytenr)
2437                         goto fail;
2438 next:
2439                 bytenr += length;
2440                 if (bytenr >= total_bytes)
2441                         break;
2442         }
2443         return 0;
2444 fail:
2445         return -1;
2446 }
2447
2448 static int do_rollback(const char *devname)
2449 {
2450         int fd = -1;
2451         int ret;
2452         int i;
2453         struct btrfs_root *root;
2454         struct btrfs_root *ext2_root;
2455         struct btrfs_root *chunk_root;
2456         struct btrfs_dir_item *dir;
2457         struct btrfs_inode_item *inode;
2458         struct btrfs_file_extent_item *fi;
2459         struct btrfs_trans_handle *trans;
2460         struct extent_buffer *leaf;
2461         struct btrfs_block_group_cache *cache1;
2462         struct btrfs_block_group_cache *cache2;
2463         struct btrfs_key key;
2464         struct btrfs_path path;
2465         struct extent_io_tree io_tree;
2466         char *buf = NULL;
2467         char *name;
2468         u64 bytenr;
2469         u64 num_bytes;
2470         u64 root_dir;
2471         u64 objectid;
2472         u64 offset;
2473         u64 start;
2474         u64 end;
2475         u64 sb_bytenr;
2476         u64 first_free;
2477         u64 total_bytes;
2478         u32 sectorsize;
2479
2480         extent_io_tree_init(&io_tree);
2481
2482         fd = open(devname, O_RDWR);
2483         if (fd < 0) {
2484                 fprintf(stderr, "unable to open %s\n", devname);
2485                 goto fail;
2486         }
2487         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2488         if (!root) {
2489                 fprintf(stderr, "unable to open ctree\n");
2490                 goto fail;
2491         }
2492         ret = may_rollback(root);
2493         if (ret < 0) {
2494                 fprintf(stderr, "unable to do rollback\n");
2495                 goto fail;
2496         }
2497
2498         sectorsize = root->sectorsize;
2499         buf = malloc(sectorsize);
2500         if (!buf) {
2501                 fprintf(stderr, "unable to allocate memory\n");
2502                 goto fail;
2503         }
2504
2505         btrfs_init_path(&path);
2506
2507         key.objectid = EXT2_IMAGE_SUBVOL_OBJECTID;
2508         key.type = BTRFS_ROOT_ITEM_KEY;
2509         key.offset = (u64)-1;
2510         ext2_root = btrfs_read_fs_root(root->fs_info, &key);
2511         if (!ext2_root || IS_ERR(ext2_root)) {
2512                 fprintf(stderr, "unable to open subvol %llu\n",
2513                         (unsigned long long)key.objectid);
2514                 goto fail;
2515         }
2516
2517         name = "image";
2518         root_dir = btrfs_root_dirid(&root->root_item);
2519         dir = btrfs_lookup_dir_item(NULL, ext2_root, &path,
2520                                    root_dir, name, strlen(name), 0);
2521         if (!dir || IS_ERR(dir)) {
2522                 fprintf(stderr, "unable to find file %s\n", name);
2523                 goto fail;
2524         }
2525         leaf = path.nodes[0];
2526         btrfs_dir_item_key_to_cpu(leaf, dir, &key);
2527         btrfs_release_path(&path);
2528
2529         objectid = key.objectid;
2530
2531         ret = btrfs_lookup_inode(NULL, ext2_root, &path, &key, 0);
2532         if (ret) {
2533                 fprintf(stderr, "unable to find inode item\n");
2534                 goto fail;
2535         }
2536         leaf = path.nodes[0];
2537         inode = btrfs_item_ptr(leaf, path.slots[0], struct btrfs_inode_item);
2538         total_bytes = btrfs_inode_size(leaf, inode);
2539         btrfs_release_path(&path);
2540
2541         key.objectid = objectid;
2542         key.offset = 0;
2543         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
2544         ret = btrfs_search_slot(NULL, ext2_root, &key, &path, 0, 0);
2545         if (ret != 0) {
2546                 fprintf(stderr, "unable to find first file extent\n");
2547                 btrfs_release_path(&path);
2548                 goto fail;
2549         }
2550
2551         /* build mapping tree for the relocated blocks */
2552         for (offset = 0; offset < total_bytes; ) {
2553                 leaf = path.nodes[0];
2554                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2555                         ret = btrfs_next_leaf(root, &path);
2556                         if (ret != 0)
2557                                 break;
2558                         continue;
2559                 }
2560
2561                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2562                 if (key.objectid != objectid || key.offset != offset ||
2563                     btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2564                         break;
2565
2566                 fi = btrfs_item_ptr(leaf, path.slots[0],
2567                                     struct btrfs_file_extent_item);
2568                 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2569                         break;
2570                 if (btrfs_file_extent_compression(leaf, fi) ||
2571                     btrfs_file_extent_encryption(leaf, fi) ||
2572                     btrfs_file_extent_other_encoding(leaf, fi))
2573                         break;
2574
2575                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2576                 /* skip holes and direct mapped extents */
2577                 if (bytenr == 0 || bytenr == offset)
2578                         goto next_extent;
2579
2580                 bytenr += btrfs_file_extent_offset(leaf, fi);
2581                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
2582
2583                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2584                 cache2 =  btrfs_lookup_block_group(root->fs_info,
2585                                                    offset + num_bytes - 1);
2586                 if (!cache1 || cache1 != cache2 ||
2587                     (!(cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
2588                      !intersect_with_sb(offset, num_bytes)))
2589                         break;
2590
2591                 set_extent_bits(&io_tree, offset, offset + num_bytes - 1,
2592                                 EXTENT_LOCKED, GFP_NOFS);
2593                 set_state_private(&io_tree, offset, bytenr);
2594 next_extent:
2595                 offset += btrfs_file_extent_num_bytes(leaf, fi);
2596                 path.slots[0]++;
2597         }
2598         btrfs_release_path(&path);
2599
2600         if (offset < total_bytes) {
2601                 fprintf(stderr, "unable to build extent mapping\n");
2602                 goto fail;
2603         }
2604
2605         first_free = BTRFS_SUPER_INFO_OFFSET + 2 * sectorsize - 1;
2606         first_free &= ~((u64)sectorsize - 1);
2607         /* backup for extent #0 should exist */
2608         if(!test_range_bit(&io_tree, 0, first_free - 1, EXTENT_LOCKED, 1)) {
2609                 fprintf(stderr, "no backup for the first extent\n");
2610                 goto fail;
2611         }
2612         /* force no allocation from system block group */
2613         root->fs_info->system_allocs = -1;
2614         trans = btrfs_start_transaction(root, 1);
2615         BUG_ON(!trans);
2616         /*
2617          * recow the whole chunk tree, this will remove all chunk tree blocks
2618          * from system block group
2619          */
2620         chunk_root = root->fs_info->chunk_root;
2621         memset(&key, 0, sizeof(key));
2622         while (1) {
2623                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2624                 if (ret < 0)
2625                         break;
2626
2627                 ret = btrfs_next_leaf(chunk_root, &path);
2628                 if (ret)
2629                         break;
2630
2631                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2632                 btrfs_release_path(&path);
2633         }
2634         btrfs_release_path(&path);
2635
2636         offset = 0;
2637         num_bytes = 0;
2638         while(1) {
2639                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2640                 if (!cache1)
2641                         break;
2642
2643                 if (cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2644                         num_bytes += btrfs_block_group_used(&cache1->item);
2645
2646                 offset = cache1->key.objectid + cache1->key.offset;
2647         }
2648         /* only extent #0 left in system block group? */
2649         if (num_bytes > first_free) {
2650                 fprintf(stderr, "unable to empty system block group\n");
2651                 goto fail;
2652         }
2653         /* create a system chunk that maps the whole device */
2654         ret = prepare_system_chunk_sb(root->fs_info->super_copy);
2655         if (ret) {
2656                 fprintf(stderr, "unable to update system chunk\n");
2657                 goto fail;
2658         }
2659
2660         ret = btrfs_commit_transaction(trans, root);
2661         BUG_ON(ret);
2662
2663         ret = close_ctree(root);
2664         if (ret) {
2665                 fprintf(stderr, "error during close_ctree %d\n", ret);
2666                 goto fail;
2667         }
2668
2669         /* zero btrfs super block mirrors */
2670         memset(buf, 0, sectorsize);
2671         for (i = 1 ; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2672                 bytenr = btrfs_sb_offset(i);
2673                 if (bytenr >= total_bytes)
2674                         break;
2675                 ret = pwrite(fd, buf, sectorsize, bytenr);
2676                 if (ret != sectorsize) {
2677                         fprintf(stderr,
2678                                 "error during zeroing supreblock %d: %d\n",
2679                                 i, ret);
2680                         goto fail;
2681                 }
2682         }
2683
2684         sb_bytenr = (u64)-1;
2685         /* copy all relocated blocks back */
2686         while(1) {
2687                 ret = find_first_extent_bit(&io_tree, 0, &start, &end,
2688                                             EXTENT_LOCKED);
2689                 if (ret)
2690                         break;
2691
2692                 ret = get_state_private(&io_tree, start, &bytenr);
2693                 BUG_ON(ret);
2694
2695                 clear_extent_bits(&io_tree, start, end, EXTENT_LOCKED,
2696                                   GFP_NOFS);
2697
2698                 while (start <= end) {
2699                         if (start == BTRFS_SUPER_INFO_OFFSET) {
2700                                 sb_bytenr = bytenr;
2701                                 goto next_sector;
2702                         }
2703                         ret = pread(fd, buf, sectorsize, bytenr);
2704                         if (ret < 0) {
2705                                 fprintf(stderr, "error during pread %d\n", ret);
2706                                 goto fail;
2707                         }
2708                         BUG_ON(ret != sectorsize);
2709                         ret = pwrite(fd, buf, sectorsize, start);
2710                         if (ret < 0) {
2711                                 fprintf(stderr, "error during pwrite %d\n", ret);
2712                                 goto fail;
2713                         }
2714                         BUG_ON(ret != sectorsize);
2715 next_sector:
2716                         start += sectorsize;
2717                         bytenr += sectorsize;
2718                 }
2719         }
2720
2721         ret = fsync(fd);
2722         if (ret) {
2723                 fprintf(stderr, "error during fsync %d\n", ret);
2724                 goto fail;
2725         }
2726         /*
2727          * finally, overwrite btrfs super block.
2728          */
2729         ret = pread(fd, buf, sectorsize, sb_bytenr);
2730         if (ret < 0) {
2731                 fprintf(stderr, "error during pread %d\n", ret);
2732                 goto fail;
2733         }
2734         BUG_ON(ret != sectorsize);
2735         ret = pwrite(fd, buf, sectorsize, BTRFS_SUPER_INFO_OFFSET);
2736         if (ret < 0) {
2737                 fprintf(stderr, "error during pwrite %d\n", ret);
2738                 goto fail;
2739         }
2740         BUG_ON(ret != sectorsize);
2741         ret = fsync(fd);
2742         if (ret) {
2743                 fprintf(stderr, "error during fsync %d\n", ret);
2744                 goto fail;
2745         }
2746
2747         close(fd);
2748         free(buf);
2749         extent_io_tree_cleanup(&io_tree);
2750         printf("rollback complete.\n");
2751         return 0;
2752
2753 fail:
2754         if (fd != -1)
2755                 close(fd);
2756         free(buf);
2757         fprintf(stderr, "rollback aborted.\n");
2758         return -1;
2759 }
2760
2761 static void print_usage(void)
2762 {
2763         printf("usage: btrfs-convert [options] device\n");
2764         printf("\t-d             disable data checksum\n");
2765         printf("\t-i             ignore xattrs and ACLs\n");
2766         printf("\t-n             disable packing of small files\n");
2767         printf("\t-r             roll back to ext2fs\n");
2768         printf("\t-l LABEL       set filesystem label\n");
2769         printf("\t-L             use label from converted fs\n");
2770         printf("\t-p             show converting progress (default)\n");
2771         printf("\t--no-progress  show only overview, not the detailed progress\n");
2772 }
2773
2774 int main(int argc, char *argv[])
2775 {
2776         int ret;
2777         int packing = 1;
2778         int noxattr = 0;
2779         int datacsum = 1;
2780         int rollback = 0;
2781         int copylabel = 0;
2782         int usage_error = 0;
2783         int progress = 1;
2784         char *file;
2785         char *fslabel = NULL;
2786
2787         while(1) {
2788                 int long_index;
2789                 enum { GETOPT_VAL_NO_PROGRESS = 256 };
2790                 static const struct option long_options[] = {
2791                         { "no-progress", no_argument, NULL, GETOPT_VAL_IEC},
2792                         { NULL, 0, NULL, 0 }
2793                 };
2794                 int c = getopt_long(argc, argv, "dinrl:Lp", long_options,
2795                                 &long_index);
2796
2797                 if (c < 0)
2798                         break;
2799                 switch(c) {
2800                         case 'd':
2801                                 datacsum = 0;
2802                                 break;
2803                         case 'i':
2804                                 noxattr = 1;
2805                                 break;
2806                         case 'n':
2807                                 packing = 0;
2808                                 break;
2809                         case 'r':
2810                                 rollback = 1;
2811                                 break;
2812                         case 'l':
2813                                 copylabel = -1;
2814                                 fslabel = strdup(optarg);
2815                                 if (strlen(fslabel) > BTRFS_LABEL_SIZE) {
2816                                         fprintf(stderr,
2817                                                 "warning: label too long, trimmed to %d bytes\n",
2818                                                 BTRFS_LABEL_SIZE);
2819                                         fslabel[BTRFS_LABEL_SIZE] = 0;
2820                                 }
2821                                 break;
2822                         case 'L':
2823                                 copylabel = 1;
2824                                 break;
2825                         case 'p':
2826                                 progress = 1;
2827                                 break;
2828                         case GETOPT_VAL_NO_PROGRESS:
2829                                 progress = 0;
2830                                 break;
2831                         default:
2832                                 print_usage();
2833                                 return 1;
2834                 }
2835         }
2836         argc = argc - optind;
2837         set_argv0(argv);
2838         if (check_argc_exact(argc, 1)) {
2839                 print_usage();
2840                 return 1;
2841         }
2842
2843         if (rollback && (!datacsum || noxattr || !packing)) {
2844                 fprintf(stderr,
2845                         "Usage error: -d, -i, -n options do not apply to rollback\n");
2846                 usage_error++;
2847         }
2848
2849         if (usage_error) {
2850                 print_usage();
2851                 return 1;
2852         }
2853
2854         file = argv[optind];
2855         ret = check_mounted(file);
2856         if (ret < 0) {
2857                 fprintf(stderr, "Could not check mount status: %s\n",
2858                         strerror(-ret));
2859                 return 1;
2860         } else if (ret) {
2861                 fprintf(stderr, "%s is mounted\n", file);
2862                 return 1;
2863         }
2864
2865         if (rollback) {
2866                 ret = do_rollback(file);
2867         } else {
2868                 ret = do_convert(file, datacsum, packing, noxattr, copylabel, fslabel, progress);
2869         }
2870         if (ret)
2871                 return 1;
2872         return 0;
2873 }