btrfs-convert.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include "kerncompat.h"
  20
  21 #include <sys/ioctl.h>
  22 #include <sys/mount.h>
  23 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <sys/types.h>
  26 #include <sys/stat.h>
  27 #include <fcntl.h>
  28 #include <unistd.h>
  29 #include <uuid/uuid.h>
  30 #include <linux/limits.h>
  31 #include <getopt.h>
  32
  33 #include "ctree.h"
  34 #include "disk-io.h"
  35 #include "volumes.h"
  36 #include "transaction.h"
  37 #include "crc32c.h"
  38 #include "utils.h"
  39 #include "task-utils.h"
  40 #include <ext2fs/ext2_fs.h>
  41 #include <ext2fs/ext2fs.h>
  42 #include <ext2fs/ext2_ext_attr.h>
  43
  44 #define INO_OFFSET (BTRFS_FIRST_FREE_OBJECTID - EXT2_ROOT_INO)
  45 #define EXT2_IMAGE_SUBVOL_OBJECTID BTRFS_FIRST_FREE_OBJECTID
  46
  47 struct task_ctx {
  48         uint32_t max_copy_inodes;
  49         uint32_t cur_copy_inodes;
  50         struct task_info *info;
  51 };
  52
  53 static void *print_copied_inodes(void *p)
  54 {
  55         struct task_ctx *priv = p;
  56         const char work_indicator[] = { '.', 'o', 'O', 'o' };
  57         uint32_t count = 0;
  58
  59         task_period_start(priv->info, 1000 /* 1s */);
  60         while (1) {
  61                 count++;
  62                 printf("copy inodes [%c] [%10d/%10d]\r",
  63                        work_indicator[count % 4], priv->cur_copy_inodes,
  64                        priv->max_copy_inodes);
  65                 fflush(stdout);
  66                 task_period_wait(priv->info);
  67         }
  68
  69         return NULL;
  70 }
  71
  72 static int after_copied_inodes(void *p)
  73 {
  74         struct task_ctx *priv = p;
  75
  76         printf("\n");
  77         task_period_stop(priv->info);
  78
  79         return 0;
  80 }
  81
  82 /*
  83  * Open Ext2fs in readonly mode, read block allocation bitmap and
  84  * inode bitmap into memory.
  85  */
  86 static int open_ext2fs(const char *name, ext2_filsys *ret_fs)
  87 {
  88         errcode_t ret;
  89         ext2_filsys ext2_fs;
  90         ext2_ino_t ino;
  91         ret = ext2fs_open(name, 0, 0, 0, unix_io_manager, &ext2_fs);
  92         if (ret) {
  93                 fprintf(stderr, "ext2fs_open: %s\n", error_message(ret));
  94                 goto fail;
  95         }
  96         ret = ext2fs_read_inode_bitmap(ext2_fs);
  97         if (ret) {
  98                 fprintf(stderr, "ext2fs_read_inode_bitmap: %s\n",
  99                         error_message(ret));
 100                 goto fail;
 101         }
 102         ret = ext2fs_read_block_bitmap(ext2_fs);
 103         if (ret) {
 104                 fprintf(stderr, "ext2fs_read_block_bitmap: %s\n",
 105                         error_message(ret));
 106                 goto fail;
 107         }
 108         /*
 109          * search each block group for a free inode. this set up
 110          * uninit block/inode bitmaps appropriately.
 111          */
 112         ino = 1;
 113         while (ino <= ext2_fs->super->s_inodes_count) {
 114                 ext2_ino_t foo;
 115                 ext2fs_new_inode(ext2_fs, ino, 0, NULL, &foo);
 116                 ino += EXT2_INODES_PER_GROUP(ext2_fs->super);
 117         }
 118
 119         *ret_fs = ext2_fs;
 120         return 0;
 121 fail:
 122         return -1;
 123 }
 124
 125 static int close_ext2fs(ext2_filsys fs)
 126 {
 127         ext2fs_close(fs);
 128         return 0;
 129 }
 130
 131 static int ext2_alloc_block(ext2_filsys fs, u64 goal, u64 *block_ret)
 132 {
 133         blk_t block;
 134
 135         if (!ext2fs_new_block(fs, goal, NULL, &block)) {
 136                 ext2fs_fast_mark_block_bitmap(fs->block_map, block);
 137                 *block_ret = block;
 138                 return 0;
 139         }
 140         return -ENOSPC;
 141 }
 142
 143 static int ext2_alloc_block_range(ext2_filsys fs, u64 goal, int num,
 144                 u64 *block_ret)
 145 {
 146         blk_t block;
 147         ext2fs_block_bitmap bitmap = fs->block_map;
 148         blk_t start = ext2fs_get_block_bitmap_start(bitmap);
 149         blk_t end = ext2fs_get_block_bitmap_end(bitmap);
 150
 151         for (block = max_t(u64, goal, start); block + num < end; block++) {
 152                 if (ext2fs_fast_test_block_bitmap_range(bitmap, block, num)) {
 153                         ext2fs_fast_mark_block_bitmap_range(bitmap, block,
 154                                         num);
 155                         *block_ret = block;
 156                         return 0;
 157                 }
 158         }
 159         return -ENOSPC;
 160 }
 161
 162 static int ext2_free_block(ext2_filsys fs, u64 block)
 163 {
 164         BUG_ON(block != (blk_t)block);
 165         ext2fs_fast_unmark_block_bitmap(fs->block_map, block);
 166         return 0;
 167 }
 168
 169 static int ext2_free_block_range(ext2_filsys fs, u64 block, int num)
 170 {
 171         BUG_ON(block != (blk_t)block);
 172         ext2fs_fast_unmark_block_bitmap_range(fs->block_map, block, num);
 173         return 0;
 174 }
 175
 176 static int cache_free_extents(struct btrfs_root *root, ext2_filsys ext2_fs)
 177
 178 {
 179         int i, ret = 0;
 180         blk_t block;
 181         u64 bytenr;
 182         u64 blocksize = ext2_fs->blocksize;
 183
 184         block = ext2_fs->super->s_first_data_block;
 185         for (; block < ext2_fs->super->s_blocks_count; block++) {
 186                 if (ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
 187                         continue;
 188                 bytenr = block * blocksize;
 189                 ret = set_extent_dirty(&root->fs_info->free_space_cache,
 190                                        bytenr, bytenr + blocksize - 1, 0);
 191                 BUG_ON(ret);
 192         }
 193
 194         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 195                 bytenr = btrfs_sb_offset(i);
 196                 bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 197                 if (bytenr >= blocksize * ext2_fs->super->s_blocks_count)
 198                         break;
 199                 clear_extent_dirty(&root->fs_info->free_space_cache, bytenr,
 200                                    bytenr + BTRFS_STRIPE_LEN - 1, 0);
 201         }
 202
 203         clear_extent_dirty(&root->fs_info->free_space_cache,
 204                            0, BTRFS_SUPER_INFO_OFFSET - 1, 0);
 205
 206         return 0;
 207 }
 208
 209 static int custom_alloc_extent(struct btrfs_root *root, u64 num_bytes,
 210                                u64 hint_byte, struct btrfs_key *ins)
 211 {
 212         u64 start;
 213         u64 end;
 214         u64 last = hint_byte;
 215         int ret;
 216         int wrapped = 0;
 217         struct btrfs_block_group_cache *cache;
 218
 219         while(1) {
 220                 ret = find_first_extent_bit(&root->fs_info->free_space_cache,
 221                                             last, &start, &end, EXTENT_DIRTY);
 222                 if (ret) {
 223                         if (wrapped++ == 0) {
 224                                 last = 0;
 225                                 continue;
 226                         } else {
 227                                 goto fail;
 228                         }
 229                 }
 230
 231                 start = max(last, start);
 232                 last = end + 1;
 233                 if (last - start < num_bytes)
 234                         continue;
 235
 236                 last = start + num_bytes;
 237                 if (test_range_bit(&root->fs_info->pinned_extents,
 238                                    start, last - 1, EXTENT_DIRTY, 0))
 239                         continue;
 240
 241                 cache = btrfs_lookup_block_group(root->fs_info, start);
 242                 BUG_ON(!cache);
 243                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM ||
 244                     last > cache->key.objectid + cache->key.offset) {
 245                         last = cache->key.objectid + cache->key.offset;
 246                         continue;
 247                 }
 248
 249                 clear_extent_dirty(&root->fs_info->free_space_cache,
 250                                    start, start + num_bytes - 1, 0);
 251
 252                 ins->objectid = start;
 253                 ins->offset = num_bytes;
 254                 ins->type = BTRFS_EXTENT_ITEM_KEY;
 255                 return 0;
 256         }
 257 fail:
 258         fprintf(stderr, "not enough free space\n");
 259         return -ENOSPC;
 260 }
 261
 262 static int intersect_with_sb(u64 bytenr, u64 num_bytes)
 263 {
 264         int i;
 265         u64 offset;
 266
 267         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 268                 offset = btrfs_sb_offset(i);
 269                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
 270
 271                 if (bytenr < offset + BTRFS_STRIPE_LEN &&
 272                     bytenr + num_bytes > offset)
 273                         return 1;
 274         }
 275         return 0;
 276 }
 277
 278 static int custom_free_extent(struct btrfs_root *root, u64 bytenr,
 279                               u64 num_bytes)
 280 {
 281         return intersect_with_sb(bytenr, num_bytes);
 282 }
 283
 284 static struct btrfs_extent_ops extent_ops = {
 285         .alloc_extent = custom_alloc_extent,
 286         .free_extent = custom_free_extent,
 287 };
 288
 289 struct dir_iterate_data {
 290         struct btrfs_trans_handle *trans;
 291         struct btrfs_root *root;
 292         struct btrfs_inode_item *inode;
 293         u64 objectid;
 294         u64 index_cnt;
 295         u64 parent;
 296         int errcode;
 297 };
 298
 299 static u8 filetype_conversion_table[EXT2_FT_MAX] = {
 300         [EXT2_FT_UNKNOWN]       = BTRFS_FT_UNKNOWN,
 301         [EXT2_FT_REG_FILE]      = BTRFS_FT_REG_FILE,
 302         [EXT2_FT_DIR]           = BTRFS_FT_DIR,
 303         [EXT2_FT_CHRDEV]        = BTRFS_FT_CHRDEV,
 304         [EXT2_FT_BLKDEV]        = BTRFS_FT_BLKDEV,
 305         [EXT2_FT_FIFO]          = BTRFS_FT_FIFO,
 306         [EXT2_FT_SOCK]          = BTRFS_FT_SOCK,
 307         [EXT2_FT_SYMLINK]       = BTRFS_FT_SYMLINK,
 308 };
 309
 310 static int dir_iterate_proc(ext2_ino_t dir, int entry,
 311                             struct ext2_dir_entry *dirent,
 312                             int offset, int blocksize,
 313                             char *buf,void *priv_data)
 314 {
 315         int ret;
 316         int file_type;
 317         u64 objectid;
 318         u64 inode_size;
 319         char dotdot[] = "..";
 320         struct btrfs_key location;
 321         struct dir_iterate_data *idata = (struct dir_iterate_data *)priv_data;
 322         int name_len;
 323
 324         name_len = dirent->name_len & 0xFF;
 325
 326         objectid = dirent->inode + INO_OFFSET;
 327         if (!strncmp(dirent->name, dotdot, name_len)) {
 328                 if (name_len == 2) {
 329                         BUG_ON(idata->parent != 0);
 330                         idata->parent = objectid;
 331                 }
 332                 return 0;
 333         }
 334         if (dirent->inode < EXT2_GOOD_OLD_FIRST_INO)
 335                 return 0;
 336
 337         location.objectid = objectid;
 338         location.offset = 0;
 339         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
 340
 341         file_type = dirent->name_len >> 8;
 342         BUG_ON(file_type > EXT2_FT_SYMLINK);
 343         ret = btrfs_insert_dir_item(idata->trans, idata->root,
 344                                     dirent->name, name_len,
 345                                     idata->objectid, &location,
 346                                     filetype_conversion_table[file_type],
 347                                     idata->index_cnt);
 348         if (ret)
 349                 goto fail;
 350         ret = btrfs_insert_inode_ref(idata->trans, idata->root,
 351                                      dirent->name, name_len,
 352                                      objectid, idata->objectid,
 353                                      idata->index_cnt);
 354         if (ret)
 355                 goto fail;
 356         idata->index_cnt++;
 357         inode_size = btrfs_stack_inode_size(idata->inode) +
 358                      name_len * 2;
 359         btrfs_set_stack_inode_size(idata->inode, inode_size);
 360         return 0;
 361 fail:
 362         idata->errcode = ret;
 363         return BLOCK_ABORT;
 364 }
 365
 366 static int create_dir_entries(struct btrfs_trans_handle *trans,
 367                               struct btrfs_root *root, u64 objectid,
 368                               struct btrfs_inode_item *btrfs_inode,
 369                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 370 {
 371         int ret;
 372         errcode_t err;
 373         struct dir_iterate_data data = {
 374                 .trans          = trans,
 375                 .root           = root,
 376                 .inode          = btrfs_inode,
 377                 .objectid       = objectid,
 378                 .index_cnt      = 2,
 379                 .parent         = 0,
 380                 .errcode        = 0,
 381         };
 382
 383         err = ext2fs_dir_iterate2(ext2_fs, ext2_ino, 0, NULL,
 384                                   dir_iterate_proc, &data);
 385         if (err)
 386                 goto error;
 387         ret = data.errcode;
 388         if (ret == 0 && data.parent == objectid) {
 389                 ret = btrfs_insert_inode_ref(trans, root, "..", 2,
 390                                              objectid, objectid, 0);
 391         }
 392         return ret;
 393 error:
 394         fprintf(stderr, "ext2fs_dir_iterate2: %s\n", error_message(err));
 395         return -1;
 396 }
 397
 398 static int read_disk_extent(struct btrfs_root *root, u64 bytenr,
 399                             u32 num_bytes, char *buffer)
 400 {
 401         int ret;
 402         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
 403
 404         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, bytenr);
 405         if (ret != num_bytes)
 406                 goto fail;
 407         ret = 0;
 408 fail:
 409         if (ret > 0)
 410                 ret = -1;
 411         return ret;
 412 }
 413
 414 static int csum_disk_extent(struct btrfs_trans_handle *trans,
 415                             struct btrfs_root *root,
 416                             u64 disk_bytenr, u64 num_bytes)
 417 {
 418         u32 blocksize = root->sectorsize;
 419         u64 offset;
 420         char *buffer;
 421         int ret = 0;
 422
 423         buffer = malloc(blocksize);
 424         if (!buffer)
 425                 return -ENOMEM;
 426         for (offset = 0; offset < num_bytes; offset += blocksize) {
 427                 ret = read_disk_extent(root, disk_bytenr + offset,
 428                                         blocksize, buffer);
 429                 if (ret)
 430                         break;
 431                 ret = btrfs_csum_file_block(trans,
 432                                             root->fs_info->csum_root,
 433                                             disk_bytenr + num_bytes,
 434                                             disk_bytenr + offset,
 435                                             buffer, blocksize);
 436                 if (ret)
 437                         break;
 438         }
 439         free(buffer);
 440         return ret;
 441 }
 442
 443 struct blk_iterate_data {
 444         struct btrfs_trans_handle *trans;
 445         struct btrfs_root *root;
 446         struct btrfs_inode_item *inode;
 447         u64 objectid;
 448         u64 first_block;
 449         u64 disk_block;
 450         u64 num_blocks;
 451         u64 boundary;
 452         int checksum;
 453         int errcode;
 454 };
 455
 456 static void init_blk_iterate_data(struct blk_iterate_data *data,
 457                                   struct btrfs_trans_handle *trans,
 458                                   struct btrfs_root *root,
 459                                   struct btrfs_inode_item *inode,
 460                                   u64 objectid, int checksum)
 461 {
 462         data->trans             = trans;
 463         data->root              = root;
 464         data->inode             = inode;
 465         data->objectid          = objectid;
 466         data->first_block       = 0;
 467         data->disk_block        = 0;
 468         data->num_blocks        = 0;
 469         data->boundary          = (u64)-1;
 470         data->checksum          = checksum;
 471         data->errcode           = 0;
 472 }
 473
 474 static int record_file_blocks(struct blk_iterate_data *data,
 475                               u64 file_block, u64 disk_block, u64 num_blocks)
 476 {
 477         int ret;
 478         struct btrfs_root *root = data->root;
 479         u64 file_pos = file_block * root->sectorsize;
 480         u64 disk_bytenr = disk_block * root->sectorsize;
 481         u64 num_bytes = num_blocks * root->sectorsize;
 482         ret = btrfs_record_file_extent(data->trans, data->root,
 483                                        data->objectid, data->inode, file_pos,
 484                                        disk_bytenr, num_bytes);
 485
 486         if (ret || !data->checksum || disk_bytenr == 0)
 487                 return ret;
 488
 489         return csum_disk_extent(data->trans, data->root, disk_bytenr,
 490                                 num_bytes);
 491 }
 492
 493 static int block_iterate_proc(u64 disk_block, u64 file_block,
 494                               struct blk_iterate_data *idata)
 495 {
 496         int ret;
 497         int sb_region;
 498         int do_barrier;
 499         struct btrfs_root *root = idata->root;
 500         struct btrfs_block_group_cache *cache;
 501         u64 bytenr = disk_block * root->sectorsize;
 502
 503         sb_region = intersect_with_sb(bytenr, root->sectorsize);
 504         do_barrier = sb_region || disk_block >= idata->boundary;
 505         if ((idata->num_blocks > 0 && do_barrier) ||
 506             (file_block > idata->first_block + idata->num_blocks) ||
 507             (disk_block != idata->disk_block + idata->num_blocks)) {
 508                 if (idata->num_blocks > 0) {
 509                         ret = record_file_blocks(idata, idata->first_block,
 510                                                  idata->disk_block,
 511                                                  idata->num_blocks);
 512                         if (ret)
 513                                 goto fail;
 514                         idata->first_block += idata->num_blocks;
 515                         idata->num_blocks = 0;
 516                 }
 517                 if (file_block > idata->first_block) {
 518                         ret = record_file_blocks(idata, idata->first_block,
 519                                         0, file_block - idata->first_block);
 520                         if (ret)
 521                                 goto fail;
 522                 }
 523
 524                 if (sb_region) {
 525                         bytenr += BTRFS_STRIPE_LEN - 1;
 526                         bytenr &= ~((u64)BTRFS_STRIPE_LEN - 1);
 527                 } else {
 528                         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
 529                         BUG_ON(!cache);
 530                         bytenr = cache->key.objectid + cache->key.offset;
 531                 }
 532
 533                 idata->first_block = file_block;
 534                 idata->disk_block = disk_block;
 535                 idata->boundary = bytenr / root->sectorsize;
 536         }
 537         idata->num_blocks++;
 538         return 0;
 539 fail:
 540         idata->errcode = ret;
 541         return BLOCK_ABORT;
 542 }
 543
 544 static int __block_iterate_proc(ext2_filsys fs, blk_t *blocknr,
 545                                 e2_blkcnt_t blockcnt, blk_t ref_block,
 546                                 int ref_offset, void *priv_data)
 547 {
 548         struct blk_iterate_data *idata;
 549         idata = (struct blk_iterate_data *)priv_data;
 550         return block_iterate_proc(*blocknr, blockcnt, idata);
 551 }
 552
 553 /*
 554  * traverse file's data blocks, record these data blocks as file extents.
 555  */
 556 static int create_file_extents(struct btrfs_trans_handle *trans,
 557                                struct btrfs_root *root, u64 objectid,
 558                                struct btrfs_inode_item *btrfs_inode,
 559                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 560                                int datacsum, int packing)
 561 {
 562         int ret;
 563         char *buffer = NULL;
 564         errcode_t err;
 565         u32 last_block;
 566         u32 sectorsize = root->sectorsize;
 567         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 568         struct blk_iterate_data data;
 569
 570         init_blk_iterate_data(&data, trans, root, btrfs_inode, objectid,
 571                               datacsum);
 572
 573         err = ext2fs_block_iterate2(ext2_fs, ext2_ino, BLOCK_FLAG_DATA_ONLY,
 574                                     NULL, __block_iterate_proc, &data);
 575         if (err)
 576                 goto error;
 577         ret = data.errcode;
 578         if (ret)
 579                 goto fail;
 580         if (packing && data.first_block == 0 && data.num_blocks > 0 &&
 581             inode_size <= BTRFS_MAX_INLINE_DATA_SIZE(root)) {
 582                 u64 num_bytes = data.num_blocks * sectorsize;
 583                 u64 disk_bytenr = data.disk_block * sectorsize;
 584                 u64 nbytes;
 585
 586                 buffer = malloc(num_bytes);
 587                 if (!buffer)
 588                         return -ENOMEM;
 589                 ret = read_disk_extent(root, disk_bytenr, num_bytes, buffer);
 590                 if (ret)
 591                         goto fail;
 592                 if (num_bytes > inode_size)
 593                         num_bytes = inode_size;
 594                 ret = btrfs_insert_inline_extent(trans, root, objectid,
 595                                                  0, buffer, num_bytes);
 596                 if (ret)
 597                         goto fail;
 598                 nbytes = btrfs_stack_inode_nbytes(btrfs_inode) + num_bytes;
 599                 btrfs_set_stack_inode_nbytes(btrfs_inode, nbytes);
 600         } else if (data.num_blocks > 0) {
 601                 ret = record_file_blocks(&data, data.first_block,
 602                                          data.disk_block, data.num_blocks);
 603                 if (ret)
 604                         goto fail;
 605         }
 606         data.first_block += data.num_blocks;
 607         last_block = (inode_size + sectorsize - 1) / sectorsize;
 608         if (last_block > data.first_block) {
 609                 ret = record_file_blocks(&data, data.first_block, 0,
 610                                          last_block - data.first_block);
 611         }
 612 fail:
 613         free(buffer);
 614         return ret;
 615 error:
 616         fprintf(stderr, "ext2fs_block_iterate2: %s\n", error_message(err));
 617         return -1;
 618 }
 619
 620 static int create_symbol_link(struct btrfs_trans_handle *trans,
 621                               struct btrfs_root *root, u64 objectid,
 622                               struct btrfs_inode_item *btrfs_inode,
 623                               ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
 624                               struct ext2_inode *ext2_inode)
 625 {
 626         int ret;
 627         char *pathname;
 628         u64 inode_size = btrfs_stack_inode_size(btrfs_inode);
 629         if (ext2fs_inode_data_blocks(ext2_fs, ext2_inode)) {
 630                 btrfs_set_stack_inode_size(btrfs_inode, inode_size + 1);
 631                 ret = create_file_extents(trans, root, objectid, btrfs_inode,
 632                                           ext2_fs, ext2_ino, 1, 1);
 633                 btrfs_set_stack_inode_size(btrfs_inode, inode_size);
 634                 return ret;
 635         }
 636
 637         pathname = (char *)&(ext2_inode->i_block[0]);
 638         BUG_ON(pathname[inode_size] != 0);
 639         ret = btrfs_insert_inline_extent(trans, root, objectid, 0,
 640                                          pathname, inode_size + 1);
 641         btrfs_set_stack_inode_nbytes(btrfs_inode, inode_size + 1);
 642         return ret;
 643 }
 644
 645 /*
 646  * Following xattr/acl related codes are based on codes in
 647  * fs/ext3/xattr.c and fs/ext3/acl.c
 648  */
 649 #define EXT2_XATTR_BHDR(ptr) ((struct ext2_ext_attr_header *)(ptr))
 650 #define EXT2_XATTR_BFIRST(ptr) \
 651         ((struct ext2_ext_attr_entry *)(EXT2_XATTR_BHDR(ptr) + 1))
 652 #define EXT2_XATTR_IHDR(inode) \
 653         ((struct ext2_ext_attr_header *) ((void *)(inode) + \
 654                 EXT2_GOOD_OLD_INODE_SIZE + (inode)->i_extra_isize))
 655 #define EXT2_XATTR_IFIRST(inode) \
 656         ((struct ext2_ext_attr_entry *) ((void *)EXT2_XATTR_IHDR(inode) + \
 657                 sizeof(EXT2_XATTR_IHDR(inode)->h_magic)))
 658
 659 static int ext2_xattr_check_names(struct ext2_ext_attr_entry *entry,
 660                                   const void *end)
 661 {
 662         struct ext2_ext_attr_entry *next;
 663
 664         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 665                 next = EXT2_EXT_ATTR_NEXT(entry);
 666                 if ((void *)next >= end)
 667                         return -EIO;
 668                 entry = next;
 669         }
 670         return 0;
 671 }
 672
 673 static int ext2_xattr_check_block(const char *buf, size_t size)
 674 {
 675         int error;
 676         struct ext2_ext_attr_header *header = EXT2_XATTR_BHDR(buf);
 677
 678         if (header->h_magic != EXT2_EXT_ATTR_MAGIC ||
 679             header->h_blocks != 1)
 680                 return -EIO;
 681         error = ext2_xattr_check_names(EXT2_XATTR_BFIRST(buf), buf + size);
 682         return error;
 683 }
 684
 685 static int ext2_xattr_check_entry(struct ext2_ext_attr_entry *entry,
 686                                   size_t size)
 687 {
 688         size_t value_size = entry->e_value_size;
 689
 690         if (entry->e_value_block != 0 || value_size > size ||
 691             entry->e_value_offs + value_size > size)
 692                 return -EIO;
 693         return 0;
 694 }
 695
 696 #define EXT2_ACL_VERSION        0x0001
 697
 698 /* 23.2.5 acl_tag_t values */
 699
 700 #define ACL_UNDEFINED_TAG       (0x00)
 701 #define ACL_USER_OBJ            (0x01)
 702 #define ACL_USER                (0x02)
 703 #define ACL_GROUP_OBJ           (0x04)
 704 #define ACL_GROUP               (0x08)
 705 #define ACL_MASK                (0x10)
 706 #define ACL_OTHER               (0x20)
 707
 708 /* 23.2.7 ACL qualifier constants */
 709
 710 #define ACL_UNDEFINED_ID        ((id_t)-1)
 711
 712 typedef struct {
 713         __le16          e_tag;
 714         __le16          e_perm;
 715         __le32          e_id;
 716 } ext2_acl_entry;
 717
 718 typedef struct {
 719         __le16          e_tag;
 720         __le16          e_perm;
 721 } ext2_acl_entry_short;
 722
 723 typedef struct {
 724         __le32          a_version;
 725 } ext2_acl_header;
 726
 727 static inline int ext2_acl_count(size_t size)
 728 {
 729         ssize_t s;
 730         size -= sizeof(ext2_acl_header);
 731         s = size - 4 * sizeof(ext2_acl_entry_short);
 732         if (s < 0) {
 733                 if (size % sizeof(ext2_acl_entry_short))
 734                         return -1;
 735                 return size / sizeof(ext2_acl_entry_short);
 736         } else {
 737                 if (s % sizeof(ext2_acl_entry))
 738                         return -1;
 739                 return s / sizeof(ext2_acl_entry) + 4;
 740         }
 741 }
 742
 743 #define ACL_EA_VERSION          0x0002
 744
 745 typedef struct {
 746         __le16          e_tag;
 747         __le16          e_perm;
 748         __le32          e_id;
 749 } acl_ea_entry;
 750
 751 typedef struct {
 752         __le32          a_version;
 753         acl_ea_entry    a_entries[0];
 754 } acl_ea_header;
 755
 756 static inline size_t acl_ea_size(int count)
 757 {
 758         return sizeof(acl_ea_header) + count * sizeof(acl_ea_entry);
 759 }
 760
 761 static int ext2_acl_to_xattr(void *dst, const void *src,
 762                              size_t dst_size, size_t src_size)
 763 {
 764         int i, count;
 765         const void *end = src + src_size;
 766         acl_ea_header *ext_acl = (acl_ea_header *)dst;
 767         acl_ea_entry *dst_entry = ext_acl->a_entries;
 768         ext2_acl_entry *src_entry;
 769
 770         if (src_size < sizeof(ext2_acl_header))
 771                 goto fail;
 772         if (((ext2_acl_header *)src)->a_version !=
 773             cpu_to_le32(EXT2_ACL_VERSION))
 774                 goto fail;
 775         src += sizeof(ext2_acl_header);
 776         count = ext2_acl_count(src_size);
 777         if (count <= 0)
 778                 goto fail;
 779
 780         BUG_ON(dst_size < acl_ea_size(count));
 781         ext_acl->a_version = cpu_to_le32(ACL_EA_VERSION);
 782         for (i = 0; i < count; i++, dst_entry++) {
 783                 src_entry = (ext2_acl_entry *)src;
 784                 if (src + sizeof(ext2_acl_entry_short) > end)
 785                         goto fail;
 786                 dst_entry->e_tag = src_entry->e_tag;
 787                 dst_entry->e_perm = src_entry->e_perm;
 788                 switch (le16_to_cpu(src_entry->e_tag)) {
 789                 case ACL_USER_OBJ:
 790                 case ACL_GROUP_OBJ:
 791                 case ACL_MASK:
 792                 case ACL_OTHER:
 793                         src += sizeof(ext2_acl_entry_short);
 794                         dst_entry->e_id = cpu_to_le32(ACL_UNDEFINED_ID);
 795                         break;
 796                 case ACL_USER:
 797                 case ACL_GROUP:
 798                         src += sizeof(ext2_acl_entry);
 799                         if (src > end)
 800                                 goto fail;
 801                         dst_entry->e_id = src_entry->e_id;
 802                         break;
 803                 default:
 804                         goto fail;
 805                 }
 806         }
 807         if (src != end)
 808                 goto fail;
 809         return 0;
 810 fail:
 811         return -EINVAL;
 812 }
 813
 814 static char *xattr_prefix_table[] = {
 815         [1] =   "user.",
 816         [2] =   "system.posix_acl_access",
 817         [3] =   "system.posix_acl_default",
 818         [4] =   "trusted.",
 819         [6] =   "security.",
 820 };
 821
 822 static int copy_single_xattr(struct btrfs_trans_handle *trans,
 823                              struct btrfs_root *root, u64 objectid,
 824                              struct ext2_ext_attr_entry *entry,
 825                              const void *data, u32 datalen)
 826 {
 827         int ret = 0;
 828         int name_len;
 829         int name_index;
 830         void *databuf = NULL;
 831         char namebuf[XATTR_NAME_MAX + 1];
 832
 833         name_index = entry->e_name_index;
 834         if (name_index >= ARRAY_SIZE(xattr_prefix_table) ||
 835             xattr_prefix_table[name_index] == NULL)
 836                 return -EOPNOTSUPP;
 837         name_len = strlen(xattr_prefix_table[name_index]) +
 838                    entry->e_name_len;
 839         if (name_len >= sizeof(namebuf))
 840                 return -ERANGE;
 841
 842         if (name_index == 2 || name_index == 3) {
 843                 size_t bufsize = acl_ea_size(ext2_acl_count(datalen));
 844                 databuf = malloc(bufsize);
 845                 if (!databuf)
 846                        return -ENOMEM;
 847                 ret = ext2_acl_to_xattr(databuf, data, bufsize, datalen);
 848                 if (ret)
 849                         goto out;
 850                 data = databuf;
 851                 datalen = bufsize;
 852         }
 853         strncpy(namebuf, xattr_prefix_table[name_index], XATTR_NAME_MAX);
 854         strncat(namebuf, EXT2_EXT_ATTR_NAME(entry), entry->e_name_len);
 855         if (name_len + datalen > BTRFS_LEAF_DATA_SIZE(root) -
 856             sizeof(struct btrfs_item) - sizeof(struct btrfs_dir_item)) {
 857                 fprintf(stderr, "skip large xattr on inode %Lu name %.*s\n",
 858                         objectid - INO_OFFSET, name_len, namebuf);
 859                 goto out;
 860         }
 861         ret = btrfs_insert_xattr_item(trans, root, namebuf, name_len,
 862                                       data, datalen, objectid);
 863 out:
 864         free(databuf);
 865         return ret;
 866 }
 867
 868 static int copy_extended_attrs(struct btrfs_trans_handle *trans,
 869                                struct btrfs_root *root, u64 objectid,
 870                                struct btrfs_inode_item *btrfs_inode,
 871                                ext2_filsys ext2_fs, ext2_ino_t ext2_ino)
 872 {
 873         int ret = 0;
 874         int inline_ea = 0;
 875         errcode_t err;
 876         u32 datalen;
 877         u32 block_size = ext2_fs->blocksize;
 878         u32 inode_size = EXT2_INODE_SIZE(ext2_fs->super);
 879         struct ext2_inode_large *ext2_inode;
 880         struct ext2_ext_attr_entry *entry;
 881         void *data;
 882         char *buffer = NULL;
 883         char inode_buf[EXT2_GOOD_OLD_INODE_SIZE];
 884
 885         if (inode_size <= EXT2_GOOD_OLD_INODE_SIZE) {
 886                 ext2_inode = (struct ext2_inode_large *)inode_buf;
 887         } else {
 888                 ext2_inode = (struct ext2_inode_large *)malloc(inode_size);
 889                 if (!ext2_inode)
 890                        return -ENOMEM;
 891         }
 892         err = ext2fs_read_inode_full(ext2_fs, ext2_ino, (void *)ext2_inode,
 893                                      inode_size);
 894         if (err) {
 895                 fprintf(stderr, "ext2fs_read_inode_full: %s\n",
 896                         error_message(err));
 897                 ret = -1;
 898                 goto out;
 899         }
 900
 901         if (ext2_ino > ext2_fs->super->s_first_ino &&
 902             inode_size > EXT2_GOOD_OLD_INODE_SIZE) {
 903                 if (EXT2_GOOD_OLD_INODE_SIZE +
 904                     ext2_inode->i_extra_isize > inode_size) {
 905                         ret = -EIO;
 906                         goto out;
 907                 }
 908                 if (ext2_inode->i_extra_isize != 0 &&
 909                     EXT2_XATTR_IHDR(ext2_inode)->h_magic ==
 910                     EXT2_EXT_ATTR_MAGIC) {
 911                         inline_ea = 1;
 912                 }
 913         }
 914         if (inline_ea) {
 915                 int total;
 916                 void *end = (void *)ext2_inode + inode_size;
 917                 entry = EXT2_XATTR_IFIRST(ext2_inode);
 918                 total = end - (void *)entry;
 919                 ret = ext2_xattr_check_names(entry, end);
 920                 if (ret)
 921                         goto out;
 922                 while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 923                         ret = ext2_xattr_check_entry(entry, total);
 924                         if (ret)
 925                                 goto out;
 926                         data = (void *)EXT2_XATTR_IFIRST(ext2_inode) +
 927                                 entry->e_value_offs;
 928                         datalen = entry->e_value_size;
 929                         ret = copy_single_xattr(trans, root, objectid,
 930                                                 entry, data, datalen);
 931                         if (ret)
 932                                 goto out;
 933                         entry = EXT2_EXT_ATTR_NEXT(entry);
 934                 }
 935         }
 936
 937         if (ext2_inode->i_file_acl == 0)
 938                 goto out;
 939
 940         buffer = malloc(block_size);
 941         if (!buffer) {
 942                 ret = -ENOMEM;
 943                 goto out;
 944         }
 945         err = ext2fs_read_ext_attr(ext2_fs, ext2_inode->i_file_acl, buffer);
 946         if (err) {
 947                 fprintf(stderr, "ext2fs_read_ext_attr: %s\n",
 948                         error_message(err));
 949                 ret = -1;
 950                 goto out;
 951         }
 952         ret = ext2_xattr_check_block(buffer, block_size);
 953         if (ret)
 954                 goto out;
 955
 956         entry = EXT2_XATTR_BFIRST(buffer);
 957         while (!EXT2_EXT_IS_LAST_ENTRY(entry)) {
 958                 ret = ext2_xattr_check_entry(entry, block_size);
 959                 if (ret)
 960                         goto out;
 961                 data = buffer + entry->e_value_offs;
 962                 datalen = entry->e_value_size;
 963                 ret = copy_single_xattr(trans, root, objectid,
 964                                         entry, data, datalen);
 965                 if (ret)
 966                         goto out;
 967                 entry = EXT2_EXT_ATTR_NEXT(entry);
 968         }
 969 out:
 970         free(buffer);
 971         if ((void *)ext2_inode != inode_buf)
 972                 free(ext2_inode);
 973         return ret;
 974 }
 975 #define MINORBITS       20
 976 #define MKDEV(ma, mi)   (((ma) << MINORBITS) | (mi))
 977
 978 static inline dev_t old_decode_dev(u16 val)
 979 {
 980         return MKDEV((val >> 8) & 255, val & 255);
 981 }
 982
 983 static inline dev_t new_decode_dev(u32 dev)
 984 {
 985         unsigned major = (dev & 0xfff00) >> 8;
 986         unsigned minor = (dev & 0xff) | ((dev >> 12) & 0xfff00);
 987         return MKDEV(major, minor);
 988 }
 989
 990 static int copy_inode_item(struct btrfs_inode_item *dst,
 991                            struct ext2_inode *src, u32 blocksize)
 992 {
 993         btrfs_set_stack_inode_generation(dst, 1);
 994         btrfs_set_stack_inode_size(dst, src->i_size);
 995         btrfs_set_stack_inode_nbytes(dst, 0);
 996         btrfs_set_stack_inode_block_group(dst, 0);
 997         btrfs_set_stack_inode_nlink(dst, src->i_links_count);
 998         btrfs_set_stack_inode_uid(dst, src->i_uid | (src->i_uid_high << 16));
 999         btrfs_set_stack_inode_gid(dst, src->i_gid | (src->i_gid_high << 16));
1000         btrfs_set_stack_inode_mode(dst, src->i_mode);
1001         btrfs_set_stack_inode_rdev(dst, 0);
1002         btrfs_set_stack_inode_flags(dst, 0);
1003         btrfs_set_stack_timespec_sec(&dst->atime, src->i_atime);
1004         btrfs_set_stack_timespec_nsec(&dst->atime, 0);
1005         btrfs_set_stack_timespec_sec(&dst->ctime, src->i_ctime);
1006         btrfs_set_stack_timespec_nsec(&dst->ctime, 0);
1007         btrfs_set_stack_timespec_sec(&dst->mtime, src->i_mtime);
1008         btrfs_set_stack_timespec_nsec(&dst->mtime, 0);
1009         btrfs_set_stack_timespec_sec(&dst->otime, 0);
1010         btrfs_set_stack_timespec_nsec(&dst->otime, 0);
1011
1012         if (S_ISDIR(src->i_mode)) {
1013                 btrfs_set_stack_inode_size(dst, 0);
1014                 btrfs_set_stack_inode_nlink(dst, 1);
1015         }
1016         if (S_ISREG(src->i_mode)) {
1017                 btrfs_set_stack_inode_size(dst, (u64)src->i_size_high << 32 |
1018                                            (u64)src->i_size);
1019         }
1020         if (!S_ISREG(src->i_mode) && !S_ISDIR(src->i_mode) &&
1021             !S_ISLNK(src->i_mode)) {
1022                 if (src->i_block[0]) {
1023                         btrfs_set_stack_inode_rdev(dst,
1024                                 old_decode_dev(src->i_block[0]));
1025                 } else {
1026                         btrfs_set_stack_inode_rdev(dst,
1027                                 new_decode_dev(src->i_block[1]));
1028                 }
1029         }
1030         return 0;
1031 }
1032
1033 /*
1034  * copy a single inode. do all the required works, such as cloning
1035  * inode item, creating file extents and creating directory entries.
1036  */
1037 static int copy_single_inode(struct btrfs_trans_handle *trans,
1038                              struct btrfs_root *root, u64 objectid,
1039                              ext2_filsys ext2_fs, ext2_ino_t ext2_ino,
1040                              struct ext2_inode *ext2_inode,
1041                              int datacsum, int packing, int noxattr)
1042 {
1043         int ret;
1044         struct btrfs_inode_item btrfs_inode;
1045
1046         if (ext2_inode->i_links_count == 0)
1047                 return 0;
1048
1049         copy_inode_item(&btrfs_inode, ext2_inode, ext2_fs->blocksize);
1050         if (!datacsum && S_ISREG(ext2_inode->i_mode)) {
1051                 u32 flags = btrfs_stack_inode_flags(&btrfs_inode) |
1052                             BTRFS_INODE_NODATASUM;
1053                 btrfs_set_stack_inode_flags(&btrfs_inode, flags);
1054         }
1055
1056         switch (ext2_inode->i_mode & S_IFMT) {
1057         case S_IFREG:
1058                 ret = create_file_extents(trans, root, objectid, &btrfs_inode,
1059                                         ext2_fs, ext2_ino, datacsum, packing);
1060                 break;
1061         case S_IFDIR:
1062                 ret = create_dir_entries(trans, root, objectid, &btrfs_inode,
1063                                          ext2_fs, ext2_ino);
1064                 break;
1065         case S_IFLNK:
1066                 ret = create_symbol_link(trans, root, objectid, &btrfs_inode,
1067                                          ext2_fs, ext2_ino, ext2_inode);
1068                 break;
1069         default:
1070                 ret = 0;
1071                 break;
1072         }
1073         if (ret)
1074                 return ret;
1075
1076         if (!noxattr) {
1077                 ret = copy_extended_attrs(trans, root, objectid, &btrfs_inode,
1078                                           ext2_fs, ext2_ino);
1079                 if (ret)
1080                         return ret;
1081         }
1082         return btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1083 }
1084
1085 static int copy_disk_extent(struct btrfs_root *root, u64 dst_bytenr,
1086                             u64 src_bytenr, u32 num_bytes)
1087 {
1088         int ret;
1089         char *buffer;
1090         struct btrfs_fs_devices *fs_devs = root->fs_info->fs_devices;
1091
1092         buffer = malloc(num_bytes);
1093         if (!buffer)
1094                 return -ENOMEM;
1095         ret = pread(fs_devs->latest_bdev, buffer, num_bytes, src_bytenr);
1096         if (ret != num_bytes)
1097                 goto fail;
1098         ret = pwrite(fs_devs->latest_bdev, buffer, num_bytes, dst_bytenr);
1099         if (ret != num_bytes)
1100                 goto fail;
1101         ret = 0;
1102 fail:
1103         free(buffer);
1104         if (ret > 0)
1105                 ret = -1;
1106         return ret;
1107 }
1108 /*
1109  * scan ext2's inode bitmap and copy all used inodes.
1110  */
1111 static int copy_inodes(struct btrfs_root *root, ext2_filsys ext2_fs,
1112                        int datacsum, int packing, int noxattr, struct task_ctx *p)
1113 {
1114         int ret;
1115         errcode_t err;
1116         ext2_inode_scan ext2_scan;
1117         struct ext2_inode ext2_inode;
1118         ext2_ino_t ext2_ino;
1119         u64 objectid;
1120         struct btrfs_trans_handle *trans;
1121
1122         trans = btrfs_start_transaction(root, 1);
1123         if (!trans)
1124                 return -ENOMEM;
1125         err = ext2fs_open_inode_scan(ext2_fs, 0, &ext2_scan);
1126         if (err) {
1127                 fprintf(stderr, "ext2fs_open_inode_scan: %s\n", error_message(err));
1128                 return -1;
1129         }
1130         while (!(err = ext2fs_get_next_inode(ext2_scan, &ext2_ino,
1131                                              &ext2_inode))) {
1132                 /* no more inodes */
1133                 if (ext2_ino == 0)
1134                         break;
1135                 /* skip special inode in ext2fs */
1136                 if (ext2_ino < EXT2_GOOD_OLD_FIRST_INO &&
1137                     ext2_ino != EXT2_ROOT_INO)
1138                         continue;
1139                 objectid = ext2_ino + INO_OFFSET;
1140                 ret = copy_single_inode(trans, root,
1141                                         objectid, ext2_fs, ext2_ino,
1142                                         &ext2_inode, datacsum, packing,
1143                                         noxattr);
1144                 p->cur_copy_inodes++;
1145                 if (ret)
1146                         return ret;
1147                 if (trans->blocks_used >= 4096) {
1148                         ret = btrfs_commit_transaction(trans, root);
1149                         BUG_ON(ret);
1150                         trans = btrfs_start_transaction(root, 1);
1151                         BUG_ON(!trans);
1152                 }
1153         }
1154         if (err) {
1155                 fprintf(stderr, "ext2fs_get_next_inode: %s\n", error_message(err));
1156                 return -1;
1157         }
1158         ret = btrfs_commit_transaction(trans, root);
1159         BUG_ON(ret);
1160
1161         return ret;
1162 }
1163
1164 /*
1165  * Construct a range of ext2fs image file.
1166  * scan block allocation bitmap, find all blocks used by the ext2fs
1167  * in this range and create file extents that point to these blocks.
1168  *
1169  * Note: Before calling the function, no file extent points to blocks
1170  *       in this range
1171  */
1172 static int create_image_file_range(struct btrfs_trans_handle *trans,
1173                                    struct btrfs_root *root, u64 objectid,
1174                                    struct btrfs_inode_item *inode,
1175                                    u64 start_byte, u64 end_byte,
1176                                    ext2_filsys ext2_fs, int datacsum)
1177 {
1178         u32 blocksize = ext2_fs->blocksize;
1179         u32 block = start_byte / blocksize;
1180         u32 last_block = (end_byte + blocksize - 1) / blocksize;
1181         int ret = 0;
1182         struct blk_iterate_data data;
1183
1184         init_blk_iterate_data(&data, trans, root, inode, objectid, datacsum);
1185         data.first_block = block;
1186
1187         for (; start_byte < end_byte; block++, start_byte += blocksize) {
1188                 if (!ext2fs_fast_test_block_bitmap(ext2_fs->block_map, block))
1189                         continue;
1190                 ret = block_iterate_proc(block, block, &data);
1191                 if (ret & BLOCK_ABORT) {
1192                         ret = data.errcode;
1193                         goto fail;
1194                 }
1195         }
1196         if (data.num_blocks > 0) {
1197                 ret = record_file_blocks(&data, data.first_block,
1198                                          data.disk_block, data.num_blocks);
1199                 if (ret)
1200                         goto fail;
1201                 data.first_block += data.num_blocks;
1202         }
1203         if (last_block > data.first_block) {
1204                 ret = record_file_blocks(&data, data.first_block, 0,
1205                                          last_block - data.first_block);
1206                 if (ret)
1207                         goto fail;
1208         }
1209 fail:
1210         return ret;
1211 }
1212 /*
1213  * Create the ext2fs image file.
1214  */
1215 static int create_ext2_image(struct btrfs_root *root, ext2_filsys ext2_fs,
1216                              const char *name, int datacsum)
1217 {
1218         int ret;
1219         struct btrfs_key key;
1220         struct btrfs_key location;
1221         struct btrfs_path path;
1222         struct btrfs_inode_item btrfs_inode;
1223         struct btrfs_inode_item *inode_item;
1224         struct extent_buffer *leaf;
1225         struct btrfs_fs_info *fs_info = root->fs_info;
1226         struct btrfs_root *extent_root = fs_info->extent_root;
1227         struct btrfs_trans_handle *trans;
1228         struct btrfs_extent_item *ei;
1229         struct btrfs_extent_inline_ref *iref;
1230         struct btrfs_extent_data_ref *dref;
1231         u64 bytenr;
1232         u64 num_bytes;
1233         u64 objectid;
1234         u64 last_byte;
1235         u64 first_free;
1236         u64 total_bytes;
1237         u64 flags = BTRFS_INODE_READONLY;
1238         u32 sectorsize = root->sectorsize;
1239
1240         total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
1241         first_free =  BTRFS_SUPER_INFO_OFFSET + sectorsize * 2 - 1;
1242         first_free &= ~((u64)sectorsize - 1);
1243         if (!datacsum)
1244                 flags |= BTRFS_INODE_NODATASUM;
1245
1246         memset(&btrfs_inode, 0, sizeof(btrfs_inode));
1247         btrfs_set_stack_inode_generation(&btrfs_inode, 1);
1248         btrfs_set_stack_inode_size(&btrfs_inode, total_bytes);
1249         btrfs_set_stack_inode_nlink(&btrfs_inode, 1);
1250         btrfs_set_stack_inode_nbytes(&btrfs_inode, 0);
1251         btrfs_set_stack_inode_mode(&btrfs_inode, S_IFREG | 0400);
1252         btrfs_set_stack_inode_flags(&btrfs_inode,  flags);
1253         btrfs_init_path(&path);
1254         trans = btrfs_start_transaction(root, 1);
1255         BUG_ON(!trans);
1256
1257         objectid = btrfs_root_dirid(&root->root_item);
1258         ret = btrfs_find_free_objectid(trans, root, objectid, &objectid);
1259         if (ret)
1260                 goto fail;
1261
1262         /*
1263          * copy blocks covered by extent #0 to new positions. extent #0 is
1264          * special, we can't rely on relocate_extents_range to relocate it.
1265          */
1266         for (last_byte = 0; last_byte < first_free; last_byte += sectorsize) {
1267                 ret = custom_alloc_extent(root, sectorsize, 0, &key);
1268                 if (ret)
1269                         goto fail;
1270                 ret = copy_disk_extent(root, key.objectid, last_byte,
1271                                        sectorsize);
1272                 if (ret)
1273                         goto fail;
1274                 ret = btrfs_record_file_extent(trans, root, objectid,
1275                                                &btrfs_inode, last_byte,
1276                                                key.objectid, sectorsize);
1277                 if (ret)
1278                         goto fail;
1279                 if (datacsum) {
1280                         ret = csum_disk_extent(trans, root, key.objectid,
1281                                                sectorsize);
1282                         if (ret)
1283                                 goto fail;
1284                 }
1285         }
1286
1287         while(1) {
1288                 key.objectid = last_byte;
1289                 key.offset = 0;
1290                 btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
1291                 ret = btrfs_search_slot(trans, fs_info->extent_root,
1292                                         &key, &path, 0, 0);
1293                 if (ret < 0)
1294                         goto fail;
1295 next:
1296                 leaf = path.nodes[0];
1297                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
1298                         ret = btrfs_next_leaf(extent_root, &path);
1299                         if (ret < 0)
1300                                 goto fail;
1301                         if (ret > 0)
1302                                 break;
1303                         leaf = path.nodes[0];
1304                 }
1305                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
1306                 if (last_byte > key.objectid ||
1307                     key.type != BTRFS_EXTENT_ITEM_KEY) {
1308                         path.slots[0]++;
1309                         goto next;
1310                 }
1311
1312                 bytenr = key.objectid;
1313                 num_bytes = key.offset;
1314                 ei = btrfs_item_ptr(leaf, path.slots[0],
1315                                     struct btrfs_extent_item);
1316                 if (!(btrfs_extent_flags(leaf, ei) & BTRFS_EXTENT_FLAG_DATA)) {
1317                         path.slots[0]++;
1318                         goto next;
1319                 }
1320
1321                 BUG_ON(btrfs_item_size_nr(leaf, path.slots[0]) != sizeof(*ei) +
1322                        btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY));
1323
1324                 iref = (struct btrfs_extent_inline_ref *)(ei + 1);
1325                 key.type = btrfs_extent_inline_ref_type(leaf, iref);
1326                 BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
1327                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1328                 if (btrfs_extent_data_ref_root(leaf, dref) !=
1329                     BTRFS_FS_TREE_OBJECTID) {
1330                         path.slots[0]++;
1331                         goto next;
1332                 }
1333
1334                 if (bytenr > last_byte) {
1335                         ret = create_image_file_range(trans, root, objectid,
1336                                                       &btrfs_inode, last_byte,
1337                                                       bytenr, ext2_fs,
1338                                                       datacsum);
1339                         if (ret)
1340                                 goto fail;
1341                 }
1342                 ret = btrfs_record_file_extent(trans, root, objectid,
1343                                                &btrfs_inode, bytenr, bytenr,
1344                                                num_bytes);
1345                 if (ret)
1346                         goto fail;
1347                 last_byte = bytenr + num_bytes;
1348                 btrfs_release_path(&path);
1349
1350                 if (trans->blocks_used >= 4096) {
1351                         ret = btrfs_commit_transaction(trans, root);
1352                         BUG_ON(ret);
1353                         trans = btrfs_start_transaction(root, 1);
1354                         BUG_ON(!trans);
1355                 }
1356         }
1357         btrfs_release_path(&path);
1358         if (total_bytes > last_byte) {
1359                 ret = create_image_file_range(trans, root, objectid,
1360                                               &btrfs_inode, last_byte,
1361                                               total_bytes, ext2_fs,
1362                                               datacsum);
1363                 if (ret)
1364                         goto fail;
1365         }
1366
1367         ret = btrfs_insert_inode(trans, root, objectid, &btrfs_inode);
1368         if (ret)
1369                 goto fail;
1370
1371         location.objectid = objectid;
1372         location.offset = 0;
1373         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1374         ret = btrfs_insert_dir_item(trans, root, name, strlen(name),
1375                                     btrfs_root_dirid(&root->root_item),
1376                                     &location, EXT2_FT_REG_FILE, objectid);
1377         if (ret)
1378                 goto fail;
1379         ret = btrfs_insert_inode_ref(trans, root, name, strlen(name),
1380                                      objectid,
1381                                      btrfs_root_dirid(&root->root_item),
1382                                      objectid);
1383         if (ret)
1384                 goto fail;
1385         location.objectid = btrfs_root_dirid(&root->root_item);
1386         location.offset = 0;
1387         btrfs_set_key_type(&location, BTRFS_INODE_ITEM_KEY);
1388         ret = btrfs_lookup_inode(trans, root, &path, &location, 1);
1389         if (ret)
1390                 goto fail;
1391         leaf = path.nodes[0];
1392         inode_item = btrfs_item_ptr(leaf, path.slots[0],
1393                                     struct btrfs_inode_item);
1394         btrfs_set_inode_size(leaf, inode_item, strlen(name) * 2 +
1395                              btrfs_inode_size(leaf, inode_item));
1396         btrfs_mark_buffer_dirty(leaf);
1397         btrfs_release_path(&path);
1398         ret = btrfs_commit_transaction(trans, root);
1399         BUG_ON(ret);
1400 fail:
1401         btrfs_release_path(&path);
1402         return ret;
1403 }
1404
1405 static struct btrfs_root * link_subvol(struct btrfs_root *root,
1406                 const char *base, u64 root_objectid)
1407 {
1408         struct btrfs_trans_handle *trans;
1409         struct btrfs_fs_info *fs_info = root->fs_info;
1410         struct btrfs_root *tree_root = fs_info->tree_root;
1411         struct btrfs_root *new_root = NULL;
1412         struct btrfs_path *path;
1413         struct btrfs_inode_item *inode_item;
1414         struct extent_buffer *leaf;
1415         struct btrfs_key key;
1416         u64 dirid = btrfs_root_dirid(&root->root_item);
1417         u64 index = 2;
1418         char buf[BTRFS_NAME_LEN + 1]; /* for snprintf null */
1419         int len;
1420         int i;
1421         int ret;
1422
1423         len = strlen(base);
1424         if (len < 1 || len > BTRFS_NAME_LEN)
1425                 return NULL;
1426
1427         path = btrfs_alloc_path();
1428         BUG_ON(!path);
1429
1430         key.objectid = dirid;
1431         key.type = BTRFS_DIR_INDEX_KEY;
1432         key.offset = (u64)-1;
1433
1434         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1435         BUG_ON(ret <= 0);
1436
1437         if (path->slots[0] > 0) {
1438                 path->slots[0]--;
1439                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1440                 if (key.objectid == dirid && key.type == BTRFS_DIR_INDEX_KEY)
1441                         index = key.offset + 1;
1442         }
1443         btrfs_release_path(path);
1444
1445         trans = btrfs_start_transaction(root, 1);
1446         BUG_ON(!trans);
1447
1448         key.objectid = dirid;
1449         key.offset = 0;
1450         key.type =  BTRFS_INODE_ITEM_KEY;
1451
1452         ret = btrfs_lookup_inode(trans, root, path, &key, 1);
1453         BUG_ON(ret);
1454         leaf = path->nodes[0];
1455         inode_item = btrfs_item_ptr(leaf, path->slots[0],
1456                                     struct btrfs_inode_item);
1457
1458         key.objectid = root_objectid;
1459         key.offset = (u64)-1;
1460         key.type = BTRFS_ROOT_ITEM_KEY;
1461
1462         memcpy(buf, base, len);
1463         for (i = 0; i < 1024; i++) {
1464                 ret = btrfs_insert_dir_item(trans, root, buf, len,
1465                                             dirid, &key, BTRFS_FT_DIR, index);
1466                 if (ret != -EEXIST)
1467                         break;
1468                 len = snprintf(buf, ARRAY_SIZE(buf), "%s%d", base, i);
1469                 if (len < 1 || len > BTRFS_NAME_LEN) {
1470                         ret = -EINVAL;
1471                         break;
1472                 }
1473         }
1474         if (ret)
1475                 goto fail;
1476
1477         btrfs_set_inode_size(leaf, inode_item, len * 2 +
1478                              btrfs_inode_size(leaf, inode_item));
1479         btrfs_mark_buffer_dirty(leaf);
1480         btrfs_release_path(path);
1481
1482         /* add the backref first */
1483         ret = btrfs_add_root_ref(trans, tree_root, root_objectid,
1484                                  BTRFS_ROOT_BACKREF_KEY,
1485                                  root->root_key.objectid,
1486                                  dirid, index, buf, len);
1487         BUG_ON(ret);
1488
1489         /* now add the forward ref */
1490         ret = btrfs_add_root_ref(trans, tree_root, root->root_key.objectid,
1491                                  BTRFS_ROOT_REF_KEY, root_objectid,
1492                                  dirid, index, buf, len);
1493
1494         ret = btrfs_commit_transaction(trans, root);
1495         BUG_ON(ret);
1496
1497         new_root = btrfs_read_fs_root(fs_info, &key);
1498         if (IS_ERR(new_root))
1499                 new_root = NULL;
1500 fail:
1501         btrfs_free_path(path);
1502         return new_root;
1503 }
1504
1505 static int create_chunk_mapping(struct btrfs_trans_handle *trans,
1506                                 struct btrfs_root *root)
1507 {
1508         struct btrfs_fs_info *info = root->fs_info;
1509         struct btrfs_root *chunk_root = info->chunk_root;
1510         struct btrfs_root *extent_root = info->extent_root;
1511         struct btrfs_device *device;
1512         struct btrfs_block_group_cache *cache;
1513         struct btrfs_dev_extent *extent;
1514         struct extent_buffer *leaf;
1515         struct btrfs_chunk chunk;
1516         struct btrfs_key key;
1517         struct btrfs_path path;
1518         u64 cur_start;
1519         u64 total_bytes;
1520         u64 chunk_objectid;
1521         int ret;
1522
1523         btrfs_init_path(&path);
1524
1525         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
1526         chunk_objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
1527
1528         BUG_ON(list_empty(&info->fs_devices->devices));
1529         device = list_entry(info->fs_devices->devices.next,
1530                             struct btrfs_device, dev_list);
1531         BUG_ON(device->devid != info->fs_devices->latest_devid);
1532
1533         /* delete device extent created by make_btrfs */
1534         key.objectid = device->devid;
1535         key.offset = 0;
1536         key.type = BTRFS_DEV_EXTENT_KEY;
1537         ret = btrfs_search_slot(trans, device->dev_root, &key, &path, -1, 1);
1538         if (ret < 0)
1539                 goto err;
1540
1541         BUG_ON(ret > 0);
1542         ret = btrfs_del_item(trans, device->dev_root, &path);
1543         if (ret)
1544                 goto err;
1545         btrfs_release_path(&path);
1546
1547         /* delete chunk item created by make_btrfs */
1548         key.objectid = chunk_objectid;
1549         key.offset = 0;
1550         key.type = BTRFS_CHUNK_ITEM_KEY;
1551         ret = btrfs_search_slot(trans, chunk_root, &key, &path, -1, 1);
1552         if (ret < 0)
1553                 goto err;
1554
1555         BUG_ON(ret > 0);
1556         ret = btrfs_del_item(trans, chunk_root, &path);
1557         if (ret)
1558                 goto err;
1559         btrfs_release_path(&path);
1560
1561         /* for each block group, create device extent and chunk item */
1562         cur_start = 0;
1563         while (cur_start < total_bytes) {
1564                 cache = btrfs_lookup_block_group(root->fs_info, cur_start);
1565                 BUG_ON(!cache);
1566
1567                 /* insert device extent */
1568                 key.objectid = device->devid;
1569                 key.offset = cache->key.objectid;
1570                 key.type = BTRFS_DEV_EXTENT_KEY;
1571                 ret = btrfs_insert_empty_item(trans, device->dev_root, &path,
1572                                               &key, sizeof(*extent));
1573                 if (ret)
1574                         goto err;
1575
1576                 leaf = path.nodes[0];
1577                 extent = btrfs_item_ptr(leaf, path.slots[0],
1578                                         struct btrfs_dev_extent);
1579
1580                 btrfs_set_dev_extent_chunk_tree(leaf, extent,
1581                                                 chunk_root->root_key.objectid);
1582                 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
1583                                                     chunk_objectid);
1584                 btrfs_set_dev_extent_chunk_offset(leaf, extent,
1585                                                   cache->key.objectid);
1586                 btrfs_set_dev_extent_length(leaf, extent, cache->key.offset);
1587                 write_extent_buffer(leaf, root->fs_info->chunk_tree_uuid,
1588                     (unsigned long)btrfs_dev_extent_chunk_tree_uuid(extent),
1589                     BTRFS_UUID_SIZE);
1590                 btrfs_mark_buffer_dirty(leaf);
1591                 btrfs_release_path(&path);
1592
1593                 /* insert chunk item */
1594                 btrfs_set_stack_chunk_length(&chunk, cache->key.offset);
1595                 btrfs_set_stack_chunk_owner(&chunk,
1596                                             extent_root->root_key.objectid);
1597                 btrfs_set_stack_chunk_stripe_len(&chunk, BTRFS_STRIPE_LEN);
1598                 btrfs_set_stack_chunk_type(&chunk, cache->flags);
1599                 btrfs_set_stack_chunk_io_align(&chunk, device->io_align);
1600                 btrfs_set_stack_chunk_io_width(&chunk, device->io_width);
1601                 btrfs_set_stack_chunk_sector_size(&chunk, device->sector_size);
1602                 btrfs_set_stack_chunk_num_stripes(&chunk, 1);
1603                 btrfs_set_stack_chunk_sub_stripes(&chunk, 0);
1604                 btrfs_set_stack_stripe_devid(&chunk.stripe, device->devid);
1605                 btrfs_set_stack_stripe_offset(&chunk.stripe,
1606                                               cache->key.objectid);
1607                 memcpy(&chunk.stripe.dev_uuid, device->uuid, BTRFS_UUID_SIZE);
1608
1609                 key.objectid = chunk_objectid;
1610                 key.offset = cache->key.objectid;
1611                 key.type = BTRFS_CHUNK_ITEM_KEY;
1612
1613                 ret = btrfs_insert_item(trans, chunk_root, &key, &chunk,
1614                                         btrfs_chunk_item_size(1));
1615                 if (ret)
1616                         goto err;
1617
1618                 cur_start = cache->key.objectid + cache->key.offset;
1619         }
1620
1621         device->bytes_used = total_bytes;
1622         ret = btrfs_update_device(trans, device);
1623 err:
1624         btrfs_release_path(&path);
1625         return ret;
1626 }
1627
1628 static int create_subvol(struct btrfs_trans_handle *trans,
1629                          struct btrfs_root *root, u64 root_objectid)
1630 {
1631         struct extent_buffer *tmp;
1632         struct btrfs_root *new_root;
1633         struct btrfs_key key;
1634         struct btrfs_root_item root_item;
1635         int ret;
1636
1637         ret = btrfs_copy_root(trans, root, root->node, &tmp,
1638                               root_objectid);
1639         BUG_ON(ret);
1640
1641         memcpy(&root_item, &root->root_item, sizeof(root_item));
1642         btrfs_set_root_bytenr(&root_item, tmp->start);
1643         btrfs_set_root_level(&root_item, btrfs_header_level(tmp));
1644         btrfs_set_root_generation(&root_item, trans->transid);
1645         free_extent_buffer(tmp);
1646
1647         key.objectid = root_objectid;
1648         key.type = BTRFS_ROOT_ITEM_KEY;
1649         key.offset = trans->transid;
1650         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
1651                                 &key, &root_item);
1652
1653         key.offset = (u64)-1;
1654         new_root = btrfs_read_fs_root(root->fs_info, &key);
1655         BUG_ON(!new_root || IS_ERR(new_root));
1656
1657         ret = btrfs_make_root_dir(trans, new_root, BTRFS_FIRST_FREE_OBJECTID);
1658         BUG_ON(ret);
1659
1660         return 0;
1661 }
1662
1663 static int init_btrfs(struct btrfs_root *root)
1664 {
1665         int ret;
1666         struct btrfs_key location;
1667         struct btrfs_trans_handle *trans;
1668         struct btrfs_fs_info *fs_info = root->fs_info;
1669         struct extent_buffer *tmp;
1670
1671         trans = btrfs_start_transaction(root, 1);
1672         BUG_ON(!trans);
1673         ret = btrfs_make_block_groups(trans, root);
1674         if (ret)
1675                 goto err;
1676         ret = btrfs_fix_block_accounting(trans, root);
1677         if (ret)
1678                 goto err;
1679         ret = create_chunk_mapping(trans, root);
1680         if (ret)
1681                 goto err;
1682         ret = btrfs_make_root_dir(trans, fs_info->tree_root,
1683                                   BTRFS_ROOT_TREE_DIR_OBJECTID);
1684         if (ret)
1685                 goto err;
1686         memcpy(&location, &root->root_key, sizeof(location));
1687         location.offset = (u64)-1;
1688         ret = btrfs_insert_dir_item(trans, fs_info->tree_root, "default", 7,
1689                                 btrfs_super_root_dir(fs_info->super_copy),
1690                                 &location, BTRFS_FT_DIR, 0);
1691         if (ret)
1692                 goto err;
1693         ret = btrfs_insert_inode_ref(trans, fs_info->tree_root, "default", 7,
1694                                 location.objectid,
1695                                 btrfs_super_root_dir(fs_info->super_copy), 0);
1696         if (ret)
1697                 goto err;
1698         btrfs_set_root_dirid(&fs_info->fs_root->root_item,
1699                              BTRFS_FIRST_FREE_OBJECTID);
1700
1701         /* subvol for ext2 image file */
1702         ret = create_subvol(trans, root, EXT2_IMAGE_SUBVOL_OBJECTID);
1703         BUG_ON(ret);
1704         /* subvol for data relocation */
1705         ret = create_subvol(trans, root, BTRFS_DATA_RELOC_TREE_OBJECTID);
1706         BUG_ON(ret);
1707
1708         extent_buffer_get(fs_info->csum_root->node);
1709         ret = __btrfs_cow_block(trans, fs_info->csum_root,
1710                                 fs_info->csum_root->node, NULL, 0, &tmp, 0, 0);
1711         BUG_ON(ret);
1712         free_extent_buffer(tmp);
1713
1714         ret = btrfs_commit_transaction(trans, root);
1715         BUG_ON(ret);
1716 err:
1717         return ret;
1718 }
1719
1720 /*
1721  * Migrate super block to its default position and zero 0 ~ 16k
1722  */
1723 static int migrate_super_block(int fd, u64 old_bytenr, u32 sectorsize)
1724 {
1725         int ret;
1726         struct extent_buffer *buf;
1727         struct btrfs_super_block *super;
1728         u32 len;
1729         u32 bytenr;
1730
1731         BUG_ON(sectorsize < sizeof(*super));
1732         buf = malloc(sizeof(*buf) + sectorsize);
1733         if (!buf)
1734                 return -ENOMEM;
1735
1736         buf->len = sectorsize;
1737         ret = pread(fd, buf->data, sectorsize, old_bytenr);
1738         if (ret != sectorsize)
1739                 goto fail;
1740
1741         super = (struct btrfs_super_block *)buf->data;
1742         BUG_ON(btrfs_super_bytenr(super) != old_bytenr);
1743         btrfs_set_super_bytenr(super, BTRFS_SUPER_INFO_OFFSET);
1744
1745         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1746         ret = pwrite(fd, buf->data, sectorsize, BTRFS_SUPER_INFO_OFFSET);
1747         if (ret != sectorsize)
1748                 goto fail;
1749
1750         ret = fsync(fd);
1751         if (ret)
1752                 goto fail;
1753
1754         memset(buf->data, 0, sectorsize);
1755         for (bytenr = 0; bytenr < BTRFS_SUPER_INFO_OFFSET; ) {
1756                 len = BTRFS_SUPER_INFO_OFFSET - bytenr;
1757                 if (len > sectorsize)
1758                         len = sectorsize;
1759                 ret = pwrite(fd, buf->data, len, bytenr);
1760                 if (ret != len) {
1761                         fprintf(stderr, "unable to zero fill device\n");
1762                         break;
1763                 }
1764                 bytenr += len;
1765         }
1766         ret = 0;
1767         fsync(fd);
1768 fail:
1769         free(buf);
1770         if (ret > 0)
1771                 ret = -1;
1772         return ret;
1773 }
1774
1775 static int prepare_system_chunk_sb(struct btrfs_super_block *super)
1776 {
1777         struct btrfs_chunk *chunk;
1778         struct btrfs_disk_key *key;
1779         u32 sectorsize = btrfs_super_sectorsize(super);
1780
1781         key = (struct btrfs_disk_key *)(super->sys_chunk_array);
1782         chunk = (struct btrfs_chunk *)(super->sys_chunk_array +
1783                                        sizeof(struct btrfs_disk_key));
1784
1785         btrfs_set_disk_key_objectid(key, BTRFS_FIRST_CHUNK_TREE_OBJECTID);
1786         btrfs_set_disk_key_type(key, BTRFS_CHUNK_ITEM_KEY);
1787         btrfs_set_disk_key_offset(key, 0);
1788
1789         btrfs_set_stack_chunk_length(chunk, btrfs_super_total_bytes(super));
1790         btrfs_set_stack_chunk_owner(chunk, BTRFS_EXTENT_TREE_OBJECTID);
1791         btrfs_set_stack_chunk_stripe_len(chunk, BTRFS_STRIPE_LEN);
1792         btrfs_set_stack_chunk_type(chunk, BTRFS_BLOCK_GROUP_SYSTEM);
1793         btrfs_set_stack_chunk_io_align(chunk, sectorsize);
1794         btrfs_set_stack_chunk_io_width(chunk, sectorsize);
1795         btrfs_set_stack_chunk_sector_size(chunk, sectorsize);
1796         btrfs_set_stack_chunk_num_stripes(chunk, 1);
1797         btrfs_set_stack_chunk_sub_stripes(chunk, 0);
1798         chunk->stripe.devid = super->dev_item.devid;
1799         btrfs_set_stack_stripe_offset(&chunk->stripe, 0);
1800         memcpy(chunk->stripe.dev_uuid, super->dev_item.uuid, BTRFS_UUID_SIZE);
1801         btrfs_set_super_sys_array_size(super, sizeof(*key) + sizeof(*chunk));
1802         return 0;
1803 }
1804
1805 static int prepare_system_chunk(int fd, u64 sb_bytenr)
1806 {
1807         int ret;
1808         struct extent_buffer *buf;
1809         struct btrfs_super_block *super;
1810
1811         BUG_ON(BTRFS_SUPER_INFO_SIZE < sizeof(*super));
1812         buf = malloc(sizeof(*buf) + BTRFS_SUPER_INFO_SIZE);
1813         if (!buf)
1814                 return -ENOMEM;
1815
1816         buf->len = BTRFS_SUPER_INFO_SIZE;
1817         ret = pread(fd, buf->data, BTRFS_SUPER_INFO_SIZE, sb_bytenr);
1818         if (ret != BTRFS_SUPER_INFO_SIZE)
1819                 goto fail;
1820
1821         super = (struct btrfs_super_block *)buf->data;
1822         BUG_ON(btrfs_super_bytenr(super) != sb_bytenr);
1823         BUG_ON(btrfs_super_num_devices(super) != 1);
1824
1825         ret = prepare_system_chunk_sb(super);
1826         if (ret)
1827                 goto fail;
1828
1829         csum_tree_block_size(buf, BTRFS_CRC32_SIZE, 0);
1830         ret = pwrite(fd, buf->data, BTRFS_SUPER_INFO_SIZE, sb_bytenr);
1831         if (ret != BTRFS_SUPER_INFO_SIZE)
1832                 goto fail;
1833
1834         ret = 0;
1835 fail:
1836         free(buf);
1837         if (ret > 0)
1838                 ret = -1;
1839         return ret;
1840 }
1841
1842 static int relocate_one_reference(struct btrfs_trans_handle *trans,
1843                                   struct btrfs_root *root,
1844                                   u64 extent_start, u64 extent_size,
1845                                   struct btrfs_key *extent_key,
1846                                   struct extent_io_tree *reloc_tree)
1847 {
1848         struct extent_buffer *leaf;
1849         struct btrfs_file_extent_item *fi;
1850         struct btrfs_key key;
1851         struct btrfs_path path;
1852         struct btrfs_inode_item inode;
1853         struct blk_iterate_data data;
1854         u64 bytenr;
1855         u64 num_bytes;
1856         u64 cur_offset;
1857         u64 new_pos;
1858         u64 nbytes;
1859         u64 sector_end;
1860         u32 sectorsize = root->sectorsize;
1861         unsigned long ptr;
1862         int datacsum;
1863         int fd;
1864         int ret;
1865
1866         btrfs_init_path(&path);
1867         ret = btrfs_search_slot(trans, root, extent_key, &path, -1, 1);
1868         if (ret)
1869                 goto fail;
1870
1871         leaf = path.nodes[0];
1872         fi = btrfs_item_ptr(leaf, path.slots[0],
1873                             struct btrfs_file_extent_item);
1874         BUG_ON(btrfs_file_extent_offset(leaf, fi) > 0);
1875         if (extent_start != btrfs_file_extent_disk_bytenr(leaf, fi) ||
1876             extent_size != btrfs_file_extent_disk_num_bytes(leaf, fi)) {
1877                 ret = 1;
1878                 goto fail;
1879         }
1880
1881         bytenr = extent_start + btrfs_file_extent_offset(leaf, fi);
1882         num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
1883
1884         ret = btrfs_del_item(trans, root, &path);
1885         if (ret)
1886                 goto fail;
1887
1888         ret = btrfs_free_extent(trans, root, extent_start, extent_size, 0,
1889                                 root->root_key.objectid,
1890                                 extent_key->objectid, extent_key->offset);
1891         if (ret)
1892                 goto fail;
1893
1894         btrfs_release_path(&path);
1895
1896         key.objectid = extent_key->objectid;
1897         key.offset = 0;
1898         key.type =  BTRFS_INODE_ITEM_KEY;
1899         ret = btrfs_lookup_inode(trans, root, &path, &key, 0);
1900         if (ret)
1901                 goto fail;
1902
1903         leaf = path.nodes[0];
1904         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1905         read_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1906         btrfs_release_path(&path);
1907
1908         BUG_ON(num_bytes & (sectorsize - 1));
1909         nbytes = btrfs_stack_inode_nbytes(&inode) - num_bytes;
1910         btrfs_set_stack_inode_nbytes(&inode, nbytes);
1911         datacsum = !(btrfs_stack_inode_flags(&inode) & BTRFS_INODE_NODATASUM);
1912
1913         init_blk_iterate_data(&data, trans, root, &inode, extent_key->objectid,
1914                               datacsum);
1915         data.first_block = extent_key->offset;
1916
1917         cur_offset = extent_key->offset;
1918         while (num_bytes > 0) {
1919                 sector_end = bytenr + sectorsize - 1;
1920                 if (test_range_bit(reloc_tree, bytenr, sector_end,
1921                                    EXTENT_LOCKED, 1)) {
1922                         ret = get_state_private(reloc_tree, bytenr, &new_pos);
1923                         BUG_ON(ret);
1924                 } else {
1925                         ret = custom_alloc_extent(root, sectorsize, 0, &key);
1926                         if (ret)
1927                                 goto fail;
1928                         new_pos = key.objectid;
1929
1930                         if (cur_offset == extent_key->offset) {
1931                                 fd = root->fs_info->fs_devices->latest_bdev;
1932                                 readahead(fd, bytenr, num_bytes);
1933                         }
1934                         ret = copy_disk_extent(root, new_pos, bytenr,
1935                                                sectorsize);
1936                         if (ret)
1937                                 goto fail;
1938                         ret = set_extent_bits(reloc_tree, bytenr, sector_end,
1939                                               EXTENT_LOCKED, GFP_NOFS);
1940                         BUG_ON(ret);
1941                         ret = set_state_private(reloc_tree, bytenr, new_pos);
1942                         BUG_ON(ret);
1943                 }
1944
1945                 ret = block_iterate_proc(new_pos / sectorsize,
1946                                          cur_offset / sectorsize, &data);
1947                 if (ret & BLOCK_ABORT) {
1948                         ret = data.errcode;
1949                         goto fail;
1950                 }
1951
1952                 cur_offset += sectorsize;
1953                 bytenr += sectorsize;
1954                 num_bytes -= sectorsize;
1955         }
1956
1957         if (data.num_blocks > 0) {
1958                 ret = record_file_blocks(&data, data.first_block,
1959                                          data.disk_block, data.num_blocks);
1960                 if (ret)
1961                         goto fail;
1962         }
1963
1964         key.objectid = extent_key->objectid;
1965         key.offset = 0;
1966         key.type =  BTRFS_INODE_ITEM_KEY;
1967         ret = btrfs_lookup_inode(trans, root, &path, &key, 1);
1968         if (ret)
1969                 goto fail;
1970
1971         leaf = path.nodes[0];
1972         ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
1973         write_extent_buffer(leaf, &inode, ptr, sizeof(inode));
1974         btrfs_mark_buffer_dirty(leaf);
1975         btrfs_release_path(&path);
1976
1977 fail:
1978         btrfs_release_path(&path);
1979         return ret;
1980 }
1981
1982 static int relocate_extents_range(struct btrfs_root *fs_root,
1983                                   struct btrfs_root *ext2_root,
1984                                   u64 start_byte, u64 end_byte)
1985 {
1986         struct btrfs_fs_info *info = fs_root->fs_info;
1987         struct btrfs_root *extent_root = info->extent_root;
1988         struct btrfs_root *cur_root = NULL;
1989         struct btrfs_trans_handle *trans;
1990         struct btrfs_extent_data_ref *dref;
1991         struct btrfs_extent_inline_ref *iref;
1992         struct btrfs_extent_item *ei;
1993         struct extent_buffer *leaf;
1994         struct btrfs_key key;
1995         struct btrfs_key extent_key;
1996         struct btrfs_path path;
1997         struct extent_io_tree reloc_tree;
1998         unsigned long ptr;
1999         unsigned long end;
2000         u64 cur_byte;
2001         u64 num_bytes;
2002         u64 ref_root;
2003         u64 num_extents;
2004         int pass = 0;
2005         int ret;
2006
2007         btrfs_init_path(&path);
2008         extent_io_tree_init(&reloc_tree);
2009
2010         key.objectid = start_byte;
2011         key.offset = 0;
2012         key.type = BTRFS_EXTENT_ITEM_KEY;
2013         ret = btrfs_search_slot(NULL, extent_root, &key, &path, 0, 0);
2014         if (ret < 0)
2015                 goto fail;
2016         if (ret > 0) {
2017                 ret = btrfs_previous_item(extent_root, &path, 0,
2018                                           BTRFS_EXTENT_ITEM_KEY);
2019                 if (ret < 0)
2020                         goto fail;
2021                 if (ret == 0) {
2022                         leaf = path.nodes[0];
2023                         btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2024                         if (key.objectid + key.offset > start_byte)
2025                                 start_byte = key.objectid;
2026                 }
2027         }
2028         btrfs_release_path(&path);
2029 again:
2030         cur_root = (pass % 2 == 0) ? ext2_root : fs_root;
2031         num_extents = 0;
2032
2033         trans = btrfs_start_transaction(cur_root, 1);
2034         BUG_ON(!trans);
2035
2036         cur_byte = start_byte;
2037         while (1) {
2038                 key.objectid = cur_byte;
2039                 key.offset = 0;
2040                 key.type = BTRFS_EXTENT_ITEM_KEY;
2041                 ret = btrfs_search_slot(trans, extent_root,
2042                                         &key, &path, 0, 0);
2043                 if (ret < 0)
2044                         goto fail;
2045 next:
2046                 leaf = path.nodes[0];
2047                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2048                         ret = btrfs_next_leaf(extent_root, &path);
2049                         if (ret < 0)
2050                                 goto fail;
2051                         if (ret > 0)
2052                                 break;
2053                         leaf = path.nodes[0];
2054                 }
2055
2056                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2057                 if (key.objectid < cur_byte ||
2058                     key.type != BTRFS_EXTENT_ITEM_KEY) {
2059                         path.slots[0]++;
2060                         goto next;
2061                 }
2062                 if (key.objectid >= end_byte)
2063                         break;
2064
2065                 num_extents++;
2066
2067                 cur_byte = key.objectid;
2068                 num_bytes = key.offset;
2069                 ei = btrfs_item_ptr(leaf, path.slots[0],
2070                                     struct btrfs_extent_item);
2071                 BUG_ON(!(btrfs_extent_flags(leaf, ei) &
2072                          BTRFS_EXTENT_FLAG_DATA));
2073
2074                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2075                 end = ptr + btrfs_item_size_nr(leaf, path.slots[0]);
2076
2077                 ptr += sizeof(struct btrfs_extent_item);
2078
2079                 while (ptr < end) {
2080                         iref = (struct btrfs_extent_inline_ref *)ptr;
2081                         key.type = btrfs_extent_inline_ref_type(leaf, iref);
2082                         BUG_ON(key.type != BTRFS_EXTENT_DATA_REF_KEY);
2083                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
2084                         ref_root = btrfs_extent_data_ref_root(leaf, dref);
2085                         extent_key.objectid =
2086                                 btrfs_extent_data_ref_objectid(leaf, dref);
2087                         extent_key.offset =
2088                                 btrfs_extent_data_ref_offset(leaf, dref);
2089                         extent_key.type = BTRFS_EXTENT_DATA_KEY;
2090                         BUG_ON(btrfs_extent_data_ref_count(leaf, dref) != 1);
2091
2092                         if (ref_root == cur_root->root_key.objectid)
2093                                 break;
2094
2095                         ptr += btrfs_extent_inline_ref_size(key.type);
2096                 }
2097
2098                 if (ptr >= end) {
2099                         path.slots[0]++;
2100                         goto next;
2101                 }
2102
2103                 ret = relocate_one_reference(trans, cur_root, cur_byte,
2104                                              num_bytes, &extent_key,
2105                                              &reloc_tree);
2106                 if (ret < 0)
2107                         goto fail;
2108
2109                 cur_byte += num_bytes;
2110                 btrfs_release_path(&path);
2111
2112                 if (trans->blocks_used >= 4096) {
2113                         ret = btrfs_commit_transaction(trans, cur_root);
2114                         BUG_ON(ret);
2115                         trans = btrfs_start_transaction(cur_root, 1);
2116                         BUG_ON(!trans);
2117                 }
2118         }
2119         btrfs_release_path(&path);
2120
2121         ret = btrfs_commit_transaction(trans, cur_root);
2122         BUG_ON(ret);
2123
2124         if (num_extents > 0 && pass++ < 16)
2125                 goto again;
2126
2127         ret = (num_extents > 0) ? -1 : 0;
2128 fail:
2129         btrfs_release_path(&path);
2130         extent_io_tree_cleanup(&reloc_tree);
2131         return ret;
2132 }
2133
2134 /*
2135  * relocate data in system chunk
2136  */
2137 static int cleanup_sys_chunk(struct btrfs_root *fs_root,
2138                              struct btrfs_root *ext2_root)
2139 {
2140         struct btrfs_block_group_cache *cache;
2141         int i, ret = 0;
2142         u64 offset = 0;
2143         u64 end_byte;
2144
2145         while(1) {
2146                 cache = btrfs_lookup_block_group(fs_root->fs_info, offset);
2147                 if (!cache)
2148                         break;
2149
2150                 end_byte = cache->key.objectid + cache->key.offset;
2151                 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2152                         ret = relocate_extents_range(fs_root, ext2_root,
2153                                                      cache->key.objectid,
2154                                                      end_byte);
2155                         if (ret)
2156                                 goto fail;
2157                 }
2158                 offset = end_byte;
2159         }
2160         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2161                 offset = btrfs_sb_offset(i);
2162                 offset &= ~((u64)BTRFS_STRIPE_LEN - 1);
2163
2164                 ret = relocate_extents_range(fs_root, ext2_root,
2165                                              offset, offset + BTRFS_STRIPE_LEN);
2166                 if (ret)
2167                         goto fail;
2168         }
2169         ret = 0;
2170 fail:
2171         return ret;
2172 }
2173
2174 static int fixup_chunk_mapping(struct btrfs_root *root)
2175 {
2176         struct btrfs_trans_handle *trans;
2177         struct btrfs_fs_info *info = root->fs_info;
2178         struct btrfs_root *chunk_root = info->chunk_root;
2179         struct extent_buffer *leaf;
2180         struct btrfs_key key;
2181         struct btrfs_path path;
2182         struct btrfs_chunk chunk;
2183         unsigned long ptr;
2184         u32 size;
2185         u64 type;
2186         int ret;
2187
2188         btrfs_init_path(&path);
2189
2190         trans = btrfs_start_transaction(root, 1);
2191         BUG_ON(!trans);
2192
2193         /*
2194          * recow the whole chunk tree. this will move all chunk tree blocks
2195          * into system block group.
2196          */
2197         memset(&key, 0, sizeof(key));
2198         while (1) {
2199                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2200                 if (ret < 0)
2201                         goto err;
2202
2203                 ret = btrfs_next_leaf(chunk_root, &path);
2204                 if (ret < 0)
2205                         goto err;
2206                 if (ret > 0)
2207                         break;
2208
2209                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2210                 btrfs_release_path(&path);
2211         }
2212         btrfs_release_path(&path);
2213
2214         /* fixup the system chunk array in super block */
2215         btrfs_set_super_sys_array_size(info->super_copy, 0);
2216
2217         key.objectid = BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2218         key.offset = 0;
2219         key.type = BTRFS_CHUNK_ITEM_KEY;
2220
2221         ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 0);
2222         if (ret < 0)
2223                 goto err;
2224         BUG_ON(ret != 0);
2225         while(1) {
2226                 leaf = path.nodes[0];
2227                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2228                         ret = btrfs_next_leaf(chunk_root, &path);
2229                         if (ret < 0)
2230                                 goto err;
2231                         if (ret > 0)
2232                                 break;
2233                         leaf = path.nodes[0];
2234                 }
2235                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2236                 if (key.type != BTRFS_CHUNK_ITEM_KEY)
2237                         goto next;
2238
2239                 ptr = btrfs_item_ptr_offset(leaf, path.slots[0]);
2240                 size = btrfs_item_size_nr(leaf, path.slots[0]);
2241                 BUG_ON(size != sizeof(chunk));
2242                 read_extent_buffer(leaf, &chunk, ptr, size);
2243                 type = btrfs_stack_chunk_type(&chunk);
2244
2245                 if (!(type & BTRFS_BLOCK_GROUP_SYSTEM))
2246                         goto next;
2247
2248                 ret = btrfs_add_system_chunk(trans, chunk_root, &key,
2249                                              &chunk, size);
2250                 if (ret)
2251                         goto err;
2252 next:
2253                 path.slots[0]++;
2254         }
2255
2256         ret = btrfs_commit_transaction(trans, root);
2257         BUG_ON(ret);
2258 err:
2259         btrfs_release_path(&path);
2260         return ret;
2261 }
2262
2263 static int do_convert(const char *devname, int datacsum, int packing, int noxattr,
2264                 u32 nodesize, int copylabel, const char *fslabel, int progress,
2265                 u64 features)
2266 {
2267         int i, ret, blocks_per_node;
2268         int fd = -1;
2269         u32 blocksize;
2270         u64 blocks[7];
2271         u64 total_bytes;
2272         u64 super_bytenr;
2273         ext2_filsys ext2_fs;
2274         struct btrfs_root *root;
2275         struct btrfs_root *ext2_root;
2276         struct task_ctx ctx;
2277         char features_buf[64];
2278
2279         ret = open_ext2fs(devname, &ext2_fs);
2280         if (ret) {
2281                 fprintf(stderr, "unable to open the Ext2fs\n");
2282                 goto fail;
2283         }
2284         blocksize = ext2_fs->blocksize;
2285         total_bytes = (u64)ext2_fs->super->s_blocks_count * blocksize;
2286         if (blocksize < 4096) {
2287                 fprintf(stderr, "block size is too small\n");
2288                 goto fail;
2289         }
2290         if (!(ext2_fs->super->s_feature_incompat &
2291               EXT2_FEATURE_INCOMPAT_FILETYPE)) {
2292                 fprintf(stderr, "filetype feature is missing\n");
2293                 goto fail;
2294         }
2295         if (btrfs_check_nodesize(nodesize, blocksize))
2296                 goto fail;
2297         blocks_per_node = nodesize / blocksize;
2298         ret = -blocks_per_node;
2299         for (i = 0; i < 7; i++) {
2300                 if (nodesize == blocksize)
2301                         ret = ext2_alloc_block(ext2_fs, 0, blocks + i);
2302                 else
2303                         ret = ext2_alloc_block_range(ext2_fs,
2304                                         ret + blocks_per_node, blocks_per_node,
2305                                         blocks + i);
2306                 if (ret) {
2307                         fprintf(stderr, "not enough free space\n");
2308                         goto fail;
2309                 }
2310                 blocks[i] *= blocksize;
2311         }
2312         super_bytenr = blocks[0];
2313         fd = open(devname, O_RDWR);
2314         if (fd < 0) {
2315                 fprintf(stderr, "unable to open %s\n", devname);
2316                 goto fail;
2317         }
2318         btrfs_parse_features_to_string(features_buf, features);
2319         if (features == BTRFS_MKFS_DEFAULT_FEATURES)
2320                 strcat(features_buf, " (default)");
2321
2322         printf("create btrfs filesystem:\n");
2323         printf("\tblocksize: %u\n", blocksize);
2324         printf("\tnodesize:  %u\n", nodesize);
2325         printf("\tfeatures:  %s\n", features_buf);
2326         ret = make_btrfs(fd, devname, ext2_fs->super->s_volume_name,
2327                          NULL, blocks, total_bytes, nodesize,
2328                          blocksize, blocksize, features);
2329         if (ret) {
2330                 fprintf(stderr, "unable to create initial ctree: %s\n",
2331                         strerror(-ret));
2332                 goto fail;
2333         }
2334         /* create a system chunk that maps the whole device */
2335         ret = prepare_system_chunk(fd, super_bytenr);
2336         if (ret) {
2337                 fprintf(stderr, "unable to update system chunk\n");
2338                 goto fail;
2339         }
2340         root = open_ctree_fd(fd, devname, super_bytenr, OPEN_CTREE_WRITES);
2341         if (!root) {
2342                 fprintf(stderr, "unable to open ctree\n");
2343                 goto fail;
2344         }
2345         ret = cache_free_extents(root, ext2_fs);
2346         if (ret) {
2347                 fprintf(stderr, "error during cache_free_extents %d\n", ret);
2348                 goto fail;
2349         }
2350         root->fs_info->extent_ops = &extent_ops;
2351         /* recover block allocation bitmap */
2352         for (i = 0; i < 7; i++) {
2353                 blocks[i] /= blocksize;
2354                 if (nodesize == blocksize)
2355                         ext2_free_block(ext2_fs, blocks[i]);
2356                 else
2357                         ext2_free_block_range(ext2_fs, blocks[i],
2358                                         blocks_per_node);
2359         }
2360         ret = init_btrfs(root);
2361         if (ret) {
2362                 fprintf(stderr, "unable to setup the root tree\n");
2363                 goto fail;
2364         }
2365         printf("creating btrfs metadata.\n");
2366         ctx.max_copy_inodes = (ext2_fs->super->s_inodes_count
2367                         - ext2_fs->super->s_free_inodes_count);
2368         ctx.cur_copy_inodes = 0;
2369
2370         if (progress) {
2371                 ctx.info = task_init(print_copied_inodes, after_copied_inodes, &ctx);
2372                 task_start(ctx.info);
2373         }
2374         ret = copy_inodes(root, ext2_fs, datacsum, packing, noxattr, &ctx);
2375         if (ret) {
2376                 fprintf(stderr, "error during copy_inodes %d\n", ret);
2377                 goto fail;
2378         }
2379         if (progress) {
2380                 task_stop(ctx.info);
2381                 task_deinit(ctx.info);
2382         }
2383         printf("creating ext2fs image file.\n");
2384         ext2_root = link_subvol(root, "ext2_saved", EXT2_IMAGE_SUBVOL_OBJECTID);
2385         if (!ext2_root) {
2386                 fprintf(stderr, "unable to create subvol\n");
2387                 goto fail;
2388         }
2389         ret = create_ext2_image(ext2_root, ext2_fs, "image", datacsum);
2390         if (ret) {
2391                 fprintf(stderr, "error during create_ext2_image %d\n", ret);
2392                 goto fail;
2393         }
2394         memset(root->fs_info->super_copy->label, 0, BTRFS_LABEL_SIZE);
2395         if (copylabel == 1) {
2396                 strncpy(root->fs_info->super_copy->label,
2397                                 ext2_fs->super->s_volume_name, 16);
2398                 fprintf(stderr, "copy label '%s'\n",
2399                                 root->fs_info->super_copy->label);
2400         } else if (copylabel == -1) {
2401                 strncpy(root->fs_info->super_copy->label, fslabel, BTRFS_LABEL_SIZE);
2402                 fprintf(stderr, "set label to '%s'\n", fslabel);
2403         }
2404
2405         printf("cleaning up system chunk.\n");
2406         ret = cleanup_sys_chunk(root, ext2_root);
2407         if (ret) {
2408                 fprintf(stderr, "error during cleanup_sys_chunk %d\n", ret);
2409                 goto fail;
2410         }
2411         ret = close_ctree(root);
2412         if (ret) {
2413                 fprintf(stderr, "error during close_ctree %d\n", ret);
2414                 goto fail;
2415         }
2416         close_ext2fs(ext2_fs);
2417
2418         /*
2419          * If this step succeed, we get a mountable btrfs. Otherwise
2420          * the ext2fs is left unchanged.
2421          */
2422         ret = migrate_super_block(fd, super_bytenr, blocksize);
2423         if (ret) {
2424                 fprintf(stderr, "unable to migrate super block\n");
2425                 goto fail;
2426         }
2427
2428         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2429         if (!root) {
2430                 fprintf(stderr, "unable to open ctree\n");
2431                 goto fail;
2432         }
2433         /* move chunk tree into system chunk. */
2434         ret = fixup_chunk_mapping(root);
2435         if (ret) {
2436                 fprintf(stderr, "error during fixup_chunk_tree\n");
2437                 goto fail;
2438         }
2439         ret = close_ctree(root);
2440         close(fd);
2441
2442         printf("conversion complete.\n");
2443         return 0;
2444 fail:
2445         if (fd != -1)
2446                 close(fd);
2447         fprintf(stderr, "conversion aborted.\n");
2448         return -1;
2449 }
2450
2451 static int may_rollback(struct btrfs_root *root)
2452 {
2453         struct btrfs_fs_info *info = root->fs_info;
2454         struct btrfs_multi_bio *multi = NULL;
2455         u64 bytenr;
2456         u64 length;
2457         u64 physical;
2458         u64 total_bytes;
2459         int num_stripes;
2460         int ret;
2461
2462         if (btrfs_super_num_devices(info->super_copy) != 1)
2463                 goto fail;
2464
2465         bytenr = BTRFS_SUPER_INFO_OFFSET;
2466         total_bytes = btrfs_super_total_bytes(root->fs_info->super_copy);
2467
2468         while (1) {
2469                 ret = btrfs_map_block(&info->mapping_tree, WRITE, bytenr,
2470                                       &length, &multi, 0, NULL);
2471                 if (ret) {
2472                         if (ret == -ENOENT) {
2473                                 /* removed block group at the tail */
2474                                 if (length == (u64)-1)
2475                                         break;
2476
2477                                 /* removed block group in the middle */
2478                                 goto next;
2479                         }
2480                         goto fail;
2481                 }
2482
2483                 num_stripes = multi->num_stripes;
2484                 physical = multi->stripes[0].physical;
2485                 kfree(multi);
2486
2487                 if (num_stripes != 1 || physical != bytenr)
2488                         goto fail;
2489 next:
2490                 bytenr += length;
2491                 if (bytenr >= total_bytes)
2492                         break;
2493         }
2494         return 0;
2495 fail:
2496         return -1;
2497 }
2498
2499 static int do_rollback(const char *devname)
2500 {
2501         int fd = -1;
2502         int ret;
2503         int i;
2504         struct btrfs_root *root;
2505         struct btrfs_root *ext2_root;
2506         struct btrfs_root *chunk_root;
2507         struct btrfs_dir_item *dir;
2508         struct btrfs_inode_item *inode;
2509         struct btrfs_file_extent_item *fi;
2510         struct btrfs_trans_handle *trans;
2511         struct extent_buffer *leaf;
2512         struct btrfs_block_group_cache *cache1;
2513         struct btrfs_block_group_cache *cache2;
2514         struct btrfs_key key;
2515         struct btrfs_path path;
2516         struct extent_io_tree io_tree;
2517         char *buf = NULL;
2518         char *name;
2519         u64 bytenr;
2520         u64 num_bytes;
2521         u64 root_dir;
2522         u64 objectid;
2523         u64 offset;
2524         u64 start;
2525         u64 end;
2526         u64 sb_bytenr;
2527         u64 first_free;
2528         u64 total_bytes;
2529         u32 sectorsize;
2530
2531         extent_io_tree_init(&io_tree);
2532
2533         fd = open(devname, O_RDWR);
2534         if (fd < 0) {
2535                 fprintf(stderr, "unable to open %s\n", devname);
2536                 goto fail;
2537         }
2538         root = open_ctree_fd(fd, devname, 0, OPEN_CTREE_WRITES);
2539         if (!root) {
2540                 fprintf(stderr, "unable to open ctree\n");
2541                 goto fail;
2542         }
2543         ret = may_rollback(root);
2544         if (ret < 0) {
2545                 fprintf(stderr, "unable to do rollback\n");
2546                 goto fail;
2547         }
2548
2549         sectorsize = root->sectorsize;
2550         buf = malloc(sectorsize);
2551         if (!buf) {
2552                 fprintf(stderr, "unable to allocate memory\n");
2553                 goto fail;
2554         }
2555
2556         btrfs_init_path(&path);
2557
2558         key.objectid = EXT2_IMAGE_SUBVOL_OBJECTID;
2559         key.type = BTRFS_ROOT_ITEM_KEY;
2560         key.offset = (u64)-1;
2561         ext2_root = btrfs_read_fs_root(root->fs_info, &key);
2562         if (!ext2_root || IS_ERR(ext2_root)) {
2563                 fprintf(stderr, "unable to open subvol %llu\n",
2564                         (unsigned long long)key.objectid);
2565                 goto fail;
2566         }
2567
2568         name = "image";
2569         root_dir = btrfs_root_dirid(&root->root_item);
2570         dir = btrfs_lookup_dir_item(NULL, ext2_root, &path,
2571                                    root_dir, name, strlen(name), 0);
2572         if (!dir || IS_ERR(dir)) {
2573                 fprintf(stderr, "unable to find file %s\n", name);
2574                 goto fail;
2575         }
2576         leaf = path.nodes[0];
2577         btrfs_dir_item_key_to_cpu(leaf, dir, &key);
2578         btrfs_release_path(&path);
2579
2580         objectid = key.objectid;
2581
2582         ret = btrfs_lookup_inode(NULL, ext2_root, &path, &key, 0);
2583         if (ret) {
2584                 fprintf(stderr, "unable to find inode item\n");
2585                 goto fail;
2586         }
2587         leaf = path.nodes[0];
2588         inode = btrfs_item_ptr(leaf, path.slots[0], struct btrfs_inode_item);
2589         total_bytes = btrfs_inode_size(leaf, inode);
2590         btrfs_release_path(&path);
2591
2592         key.objectid = objectid;
2593         key.offset = 0;
2594         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
2595         ret = btrfs_search_slot(NULL, ext2_root, &key, &path, 0, 0);
2596         if (ret != 0) {
2597                 fprintf(stderr, "unable to find first file extent\n");
2598                 btrfs_release_path(&path);
2599                 goto fail;
2600         }
2601
2602         /* build mapping tree for the relocated blocks */
2603         for (offset = 0; offset < total_bytes; ) {
2604                 leaf = path.nodes[0];
2605                 if (path.slots[0] >= btrfs_header_nritems(leaf)) {
2606                         ret = btrfs_next_leaf(root, &path);
2607                         if (ret != 0)
2608                                 break;
2609                         continue;
2610                 }
2611
2612                 btrfs_item_key_to_cpu(leaf, &key, path.slots[0]);
2613                 if (key.objectid != objectid || key.offset != offset ||
2614                     btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2615                         break;
2616
2617                 fi = btrfs_item_ptr(leaf, path.slots[0],
2618                                     struct btrfs_file_extent_item);
2619                 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2620                         break;
2621                 if (btrfs_file_extent_compression(leaf, fi) ||
2622                     btrfs_file_extent_encryption(leaf, fi) ||
2623                     btrfs_file_extent_other_encoding(leaf, fi))
2624                         break;
2625
2626                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
2627                 /* skip holes and direct mapped extents */
2628                 if (bytenr == 0 || bytenr == offset)
2629                         goto next_extent;
2630
2631                 bytenr += btrfs_file_extent_offset(leaf, fi);
2632                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
2633
2634                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2635                 cache2 =  btrfs_lookup_block_group(root->fs_info,
2636                                                    offset + num_bytes - 1);
2637                 if (!cache1 || cache1 != cache2 ||
2638                     (!(cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
2639                      !intersect_with_sb(offset, num_bytes)))
2640                         break;
2641
2642                 set_extent_bits(&io_tree, offset, offset + num_bytes - 1,
2643                                 EXTENT_LOCKED, GFP_NOFS);
2644                 set_state_private(&io_tree, offset, bytenr);
2645 next_extent:
2646                 offset += btrfs_file_extent_num_bytes(leaf, fi);
2647                 path.slots[0]++;
2648         }
2649         btrfs_release_path(&path);
2650
2651         if (offset < total_bytes) {
2652                 fprintf(stderr, "unable to build extent mapping\n");
2653                 goto fail;
2654         }
2655
2656         first_free = BTRFS_SUPER_INFO_OFFSET + 2 * sectorsize - 1;
2657         first_free &= ~((u64)sectorsize - 1);
2658         /* backup for extent #0 should exist */
2659         if(!test_range_bit(&io_tree, 0, first_free - 1, EXTENT_LOCKED, 1)) {
2660                 fprintf(stderr, "no backup for the first extent\n");
2661                 goto fail;
2662         }
2663         /* force no allocation from system block group */
2664         root->fs_info->system_allocs = -1;
2665         trans = btrfs_start_transaction(root, 1);
2666         BUG_ON(!trans);
2667         /*
2668          * recow the whole chunk tree, this will remove all chunk tree blocks
2669          * from system block group
2670          */
2671         chunk_root = root->fs_info->chunk_root;
2672         memset(&key, 0, sizeof(key));
2673         while (1) {
2674                 ret = btrfs_search_slot(trans, chunk_root, &key, &path, 0, 1);
2675                 if (ret < 0)
2676                         break;
2677
2678                 ret = btrfs_next_leaf(chunk_root, &path);
2679                 if (ret)
2680                         break;
2681
2682                 btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
2683                 btrfs_release_path(&path);
2684         }
2685         btrfs_release_path(&path);
2686
2687         offset = 0;
2688         num_bytes = 0;
2689         while(1) {
2690                 cache1 = btrfs_lookup_block_group(root->fs_info, offset);
2691                 if (!cache1)
2692                         break;
2693
2694                 if (cache1->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2695                         num_bytes += btrfs_block_group_used(&cache1->item);
2696
2697                 offset = cache1->key.objectid + cache1->key.offset;
2698         }
2699         /* only extent #0 left in system block group? */
2700         if (num_bytes > first_free) {
2701                 fprintf(stderr, "unable to empty system block group\n");
2702                 goto fail;
2703         }
2704         /* create a system chunk that maps the whole device */
2705         ret = prepare_system_chunk_sb(root->fs_info->super_copy);
2706         if (ret) {
2707                 fprintf(stderr, "unable to update system chunk\n");
2708                 goto fail;
2709         }
2710
2711         ret = btrfs_commit_transaction(trans, root);
2712         BUG_ON(ret);
2713
2714         ret = close_ctree(root);
2715         if (ret) {
2716                 fprintf(stderr, "error during close_ctree %d\n", ret);
2717                 goto fail;
2718         }
2719
2720         /* zero btrfs super block mirrors */
2721         memset(buf, 0, sectorsize);
2722         for (i = 1 ; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2723                 bytenr = btrfs_sb_offset(i);
2724                 if (bytenr >= total_bytes)
2725                         break;
2726                 ret = pwrite(fd, buf, sectorsize, bytenr);
2727                 if (ret != sectorsize) {
2728                         fprintf(stderr,
2729                                 "error during zeroing supreblock %d: %d\n",
2730                                 i, ret);
2731                         goto fail;
2732                 }
2733         }
2734
2735         sb_bytenr = (u64)-1;
2736         /* copy all relocated blocks back */
2737         while(1) {
2738                 ret = find_first_extent_bit(&io_tree, 0, &start, &end,
2739                                             EXTENT_LOCKED);
2740                 if (ret)
2741                         break;
2742
2743                 ret = get_state_private(&io_tree, start, &bytenr);
2744                 BUG_ON(ret);
2745
2746                 clear_extent_bits(&io_tree, start, end, EXTENT_LOCKED,
2747                                   GFP_NOFS);
2748
2749                 while (start <= end) {
2750                         if (start == BTRFS_SUPER_INFO_OFFSET) {
2751                                 sb_bytenr = bytenr;
2752                                 goto next_sector;
2753                         }
2754                         ret = pread(fd, buf, sectorsize, bytenr);
2755                         if (ret < 0) {
2756                                 fprintf(stderr, "error during pread %d\n", ret);
2757                                 goto fail;
2758                         }
2759                         BUG_ON(ret != sectorsize);
2760                         ret = pwrite(fd, buf, sectorsize, start);
2761                         if (ret < 0) {
2762                                 fprintf(stderr, "error during pwrite %d\n", ret);
2763                                 goto fail;
2764                         }
2765                         BUG_ON(ret != sectorsize);
2766 next_sector:
2767                         start += sectorsize;
2768                         bytenr += sectorsize;
2769                 }
2770         }
2771
2772         ret = fsync(fd);
2773         if (ret) {
2774                 fprintf(stderr, "error during fsync %d\n", ret);
2775                 goto fail;
2776         }
2777         /*
2778          * finally, overwrite btrfs super block.
2779          */
2780         ret = pread(fd, buf, sectorsize, sb_bytenr);
2781         if (ret < 0) {
2782                 fprintf(stderr, "error during pread %d\n", ret);
2783                 goto fail;
2784         }
2785         BUG_ON(ret != sectorsize);
2786         ret = pwrite(fd, buf, sectorsize, BTRFS_SUPER_INFO_OFFSET);
2787         if (ret < 0) {
2788                 fprintf(stderr, "error during pwrite %d\n", ret);
2789                 goto fail;
2790         }
2791         BUG_ON(ret != sectorsize);
2792         ret = fsync(fd);
2793         if (ret) {
2794                 fprintf(stderr, "error during fsync %d\n", ret);
2795                 goto fail;
2796         }
2797
2798         close(fd);
2799         free(buf);
2800         extent_io_tree_cleanup(&io_tree);
2801         printf("rollback complete.\n");
2802         return 0;
2803
2804 fail:
2805         if (fd != -1)
2806                 close(fd);
2807         free(buf);
2808         fprintf(stderr, "rollback aborted.\n");
2809         return -1;
2810 }
2811
2812 static void print_usage(void)
2813 {
2814         printf("usage: btrfs-convert [options] device\n");
2815         printf("options:\n");
2816         printf("\t-d|--no-datasum        disable data checksum, sets NODATASUM\n");
2817         printf("\t-i|--no-xattr          ignore xattrs and ACLs\n");
2818         printf("\t-n|--no-inline         disable inlining of small files to metadata\n");
2819         printf("\t-N|--nodesize SIZE     set filesystem metadata nodesize\n");
2820         printf("\t-r|--rollback          roll back to ext2fs\n");
2821         printf("\t-l|--label LABEL       set filesystem label\n");
2822         printf("\t-L|--copy-label        use label from converted filesystem\n");
2823         printf("\t-p|--progress          show converting progress (default)\n");
2824         printf("\t-O|--features LIST     comma separated list of filesystem features\n");
2825         printf("\t--no-progress          show only overview, not the detailed progress\n");
2826 }
2827
2828 int main(int argc, char *argv[])
2829 {
2830         int ret;
2831         int packing = 1;
2832         int noxattr = 0;
2833         int datacsum = 1;
2834         u32 nodesize = max_t(u32, sysconf(_SC_PAGESIZE),
2835                         BTRFS_MKFS_DEFAULT_NODE_SIZE);
2836         int rollback = 0;
2837         int copylabel = 0;
2838         int usage_error = 0;
2839         int progress = 1;
2840         char *file;
2841         char *fslabel = NULL;
2842         u64 features = BTRFS_MKFS_DEFAULT_FEATURES;
2843
2844         while(1) {
2845                 enum { GETOPT_VAL_NO_PROGRESS = 256 };
2846                 static const struct option long_options[] = {
2847                         { "no-progress", no_argument, NULL,
2848                                 GETOPT_VAL_NO_PROGRESS },
2849                         { "no-datasum", no_argument, NULL, 'd' },
2850                         { "no-inline", no_argument, NULL, 'n' },
2851                         { "no-xattr", no_argument, NULL, 'i' },
2852                         { "rollback", no_argument, NULL, 'r' },
2853                         { "features", required_argument, NULL, 'O' },
2854                         { "progress", no_argument, NULL, 'p' },
2855                         { "label", required_argument, NULL, 'l' },
2856                         { "copy-label", no_argument, NULL, 'L' },
2857                         { "nodesize", required_argument, NULL, 'N' },
2858                         { NULL, 0, NULL, 0 }
2859                 };
2860                 int c = getopt_long(argc, argv, "dinN:rl:LpO:", long_options, NULL);
2861
2862                 if (c < 0)
2863                         break;
2864                 switch(c) {
2865                         case 'd':
2866                                 datacsum = 0;
2867                                 break;
2868                         case 'i':
2869                                 noxattr = 1;
2870                                 break;
2871                         case 'n':
2872                                 packing = 0;
2873                                 break;
2874                         case 'N':
2875                                 nodesize = parse_size(optarg);
2876                                 break;
2877                         case 'r':
2878                                 rollback = 1;
2879                                 break;
2880                         case 'l':
2881                                 copylabel = -1;
2882                                 fslabel = strdup(optarg);
2883                                 if (strlen(fslabel) > BTRFS_LABEL_SIZE) {
2884                                         fprintf(stderr,
2885                                                 "warning: label too long, trimmed to %d bytes\n",
2886                                                 BTRFS_LABEL_SIZE);
2887                                         fslabel[BTRFS_LABEL_SIZE] = 0;
2888                                 }
2889                                 break;
2890                         case 'L':
2891                                 copylabel = 1;
2892                                 break;
2893                         case 'p':
2894                                 progress = 1;
2895                                 break;
2896                         case 'O': {
2897                                 char *orig = strdup(optarg);
2898                                 char *tmp = orig;
2899
2900                                 tmp = btrfs_parse_fs_features(tmp, &features);
2901                                 if (tmp) {
2902                                         fprintf(stderr,
2903                                                 "Unrecognized filesystem feature '%s'\n",
2904                                                         tmp);
2905                                         free(orig);
2906                                         exit(1);
2907                                 }
2908                                 free(orig);
2909                                 if (features & BTRFS_FEATURE_LIST_ALL) {
2910                                         btrfs_list_all_fs_features(
2911                                                 ~BTRFS_CONVERT_ALLOWED_FEATURES);
2912                                         exit(0);
2913                                 }
2914                                 if (features & ~BTRFS_CONVERT_ALLOWED_FEATURES) {
2915                                         char buf[64];
2916
2917                                         btrfs_parse_features_to_string(buf,
2918                                                 features & ~BTRFS_CONVERT_ALLOWED_FEATURES);
2919                                         fprintf(stderr,
2920                                                 "ERROR: features not allowed for convert: %s\n",
2921                                                 buf);
2922                                         exit(1);
2923                                 }
2924
2925                                 break;
2926                                 }
2927                         case GETOPT_VAL_NO_PROGRESS:
2928                                 progress = 0;
2929                                 break;
2930                         default:
2931                                 print_usage();
2932                                 return 1;
2933                 }
2934         }
2935         argc = argc - optind;
2936         set_argv0(argv);
2937         if (check_argc_exact(argc, 1)) {
2938                 print_usage();
2939                 return 1;
2940         }
2941
2942         if (rollback && (!datacsum || noxattr || !packing)) {
2943                 fprintf(stderr,
2944                         "Usage error: -d, -i, -n options do not apply to rollback\n");
2945                 usage_error++;
2946         }
2947
2948         if (usage_error) {
2949                 print_usage();
2950                 return 1;
2951         }
2952
2953         file = argv[optind];
2954         ret = check_mounted(file);
2955         if (ret < 0) {
2956                 fprintf(stderr, "Could not check mount status: %s\n",
2957                         strerror(-ret));
2958                 return 1;
2959         } else if (ret) {
2960                 fprintf(stderr, "%s is mounted\n", file);
2961                 return 1;
2962         }
2963
2964         if (rollback) {
2965                 ret = do_rollback(file);
2966         } else {
2967                 ret = do_convert(file, datacsum, packing, noxattr, nodesize,
2968                                 copylabel, fslabel, progress, features);
2969         }
2970         if (ret)
2971                 return 1;
2972         return 0;
2973 }