Pileus Git - ~andy/linux/blob - fs/ext3/resize.c

   1 /*
   2  *  linux/fs/ext3/resize.c
   3  *
   4  * Support for resizing an ext3 filesystem while it is mounted.
   5  *
   6  * Copyright (C) 2001, 2002 Andreas Dilger <adilger@clusterfs.com>
   7  *
   8  * This could probably be made into a module, because it is not often in use.
   9  */
  10
  11 #include <linux/config.h>
  12
  13 #define EXT3FS_DEBUG
  14
  15 #include <linux/sched.h>
  16 #include <linux/smp_lock.h>
  17 #include <linux/ext3_jbd.h>
  18
  19 #include <linux/errno.h>
  20 #include <linux/slab.h>
  21
  22
  23 #define outside(b, first, last) ((b) < (first) || (b) >= (last))
  24 #define inside(b, first, last)  ((b) >= (first) && (b) < (last))
  25
  26 static int verify_group_input(struct super_block *sb,
  27                               struct ext3_new_group_data *input)
  28 {
  29         struct ext3_sb_info *sbi = EXT3_SB(sb);
  30         struct ext3_super_block *es = sbi->s_es;
  31         ext3_fsblk_t start = le32_to_cpu(es->s_blocks_count);
  32         ext3_fsblk_t end = start + input->blocks_count;
  33         unsigned group = input->group;
  34         ext3_fsblk_t itend = input->inode_table + sbi->s_itb_per_group;
  35         unsigned overhead = ext3_bg_has_super(sb, group) ?
  36                 (1 + ext3_bg_num_gdb(sb, group) +
  37                  le16_to_cpu(es->s_reserved_gdt_blocks)) : 0;
  38         ext3_fsblk_t metaend = start + overhead;
  39         struct buffer_head *bh = NULL;
  40         ext3_grpblk_t free_blocks_count;
  41         int err = -EINVAL;
  42
  43         input->free_blocks_count = free_blocks_count =
  44                 input->blocks_count - 2 - overhead - sbi->s_itb_per_group;
  45
  46         if (test_opt(sb, DEBUG))
  47                 printk(KERN_DEBUG "EXT3-fs: adding %s group %u: %u blocks "
  48                        "(%d free, %u reserved)\n",
  49                        ext3_bg_has_super(sb, input->group) ? "normal" :
  50                        "no-super", input->group, input->blocks_count,
  51                        free_blocks_count, input->reserved_blocks);
  52
  53         if (group != sbi->s_groups_count)
  54                 ext3_warning(sb, __FUNCTION__,
  55                              "Cannot add at group %u (only %lu groups)",
  56                              input->group, sbi->s_groups_count);
  57         else if ((start - le32_to_cpu(es->s_first_data_block)) %
  58                  EXT3_BLOCKS_PER_GROUP(sb))
  59                 ext3_warning(sb, __FUNCTION__, "Last group not full");
  60         else if (input->reserved_blocks > input->blocks_count / 5)
  61                 ext3_warning(sb, __FUNCTION__, "Reserved blocks too high (%u)",
  62                              input->reserved_blocks);
  63         else if (free_blocks_count < 0)
  64                 ext3_warning(sb, __FUNCTION__, "Bad blocks count %u",
  65                              input->blocks_count);
  66         else if (!(bh = sb_bread(sb, end - 1)))
  67                 ext3_warning(sb, __FUNCTION__,
  68                              "Cannot read last block ("E3FSBLK")",
  69                              end - 1);
  70         else if (outside(input->block_bitmap, start, end))
  71                 ext3_warning(sb, __FUNCTION__,
  72                              "Block bitmap not in group (block %u)",
  73                              input->block_bitmap);
  74         else if (outside(input->inode_bitmap, start, end))
  75                 ext3_warning(sb, __FUNCTION__,
  76                              "Inode bitmap not in group (block %u)",
  77                              input->inode_bitmap);
  78         else if (outside(input->inode_table, start, end) ||
  79                  outside(itend - 1, start, end))
  80                 ext3_warning(sb, __FUNCTION__,
  81                              "Inode table not in group (blocks %u-"E3FSBLK")",
  82                              input->inode_table, itend - 1);
  83         else if (input->inode_bitmap == input->block_bitmap)
  84                 ext3_warning(sb, __FUNCTION__,
  85                              "Block bitmap same as inode bitmap (%u)",
  86                              input->block_bitmap);
  87         else if (inside(input->block_bitmap, input->inode_table, itend))
  88                 ext3_warning(sb, __FUNCTION__,
  89                              "Block bitmap (%u) in inode table (%u-"E3FSBLK")",
  90                              input->block_bitmap, input->inode_table, itend-1);
  91         else if (inside(input->inode_bitmap, input->inode_table, itend))
  92                 ext3_warning(sb, __FUNCTION__,
  93                              "Inode bitmap (%u) in inode table (%u-"E3FSBLK")",
  94                              input->inode_bitmap, input->inode_table, itend-1);
  95         else if (inside(input->block_bitmap, start, metaend))
  96                 ext3_warning(sb, __FUNCTION__,
  97                              "Block bitmap (%u) in GDT table"
  98                              " ("E3FSBLK"-"E3FSBLK")",
  99                              input->block_bitmap, start, metaend - 1);
 100         else if (inside(input->inode_bitmap, start, metaend))
 101                 ext3_warning(sb, __FUNCTION__,
 102                              "Inode bitmap (%u) in GDT table"
 103                              " ("E3FSBLK"-"E3FSBLK")",
 104                              input->inode_bitmap, start, metaend - 1);
 105         else if (inside(input->inode_table, start, metaend) ||
 106                  inside(itend - 1, start, metaend))
 107                 ext3_warning(sb, __FUNCTION__,
 108                              "Inode table (%u-"E3FSBLK") overlaps"
 109                              "GDT table ("E3FSBLK"-"E3FSBLK")",
 110                              input->inode_table, itend - 1, start, metaend - 1);
 111         else
 112                 err = 0;
 113         brelse(bh);
 114
 115         return err;
 116 }
 117
 118 static struct buffer_head *bclean(handle_t *handle, struct super_block *sb,
 119                                   unsigned long blk)
 120 {
 121         struct buffer_head *bh;
 122         int err;
 123
 124         bh = sb_getblk(sb, blk);
 125         if (!bh)
 126                 return ERR_PTR(-EIO);
 127         if ((err = ext3_journal_get_write_access(handle, bh))) {
 128                 brelse(bh);
 129                 bh = ERR_PTR(err);
 130         } else {
 131                 lock_buffer(bh);
 132                 memset(bh->b_data, 0, sb->s_blocksize);
 133                 set_buffer_uptodate(bh);
 134                 unlock_buffer(bh);
 135         }
 136
 137         return bh;
 138 }
 139
 140 /*
 141  * To avoid calling the atomic setbit hundreds or thousands of times, we only
 142  * need to use it within a single byte (to ensure we get endianness right).
 143  * We can use memset for the rest of the bitmap as there are no other users.
 144  */
 145 static void mark_bitmap_end(int start_bit, int end_bit, char *bitmap)
 146 {
 147         int i;
 148
 149         if (start_bit >= end_bit)
 150                 return;
 151
 152         ext3_debug("mark end bits +%d through +%d used\n", start_bit, end_bit);
 153         for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++)
 154                 ext3_set_bit(i, bitmap);
 155         if (i < end_bit)
 156                 memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3);
 157 }
 158
 159 /*
 160  * Set up the block and inode bitmaps, and the inode table for the new group.
 161  * This doesn't need to be part of the main transaction, since we are only
 162  * changing blocks outside the actual filesystem.  We still do journaling to
 163  * ensure the recovery is correct in case of a failure just after resize.
 164  * If any part of this fails, we simply abort the resize.
 165  */
 166 static int setup_new_group_blocks(struct super_block *sb,
 167                                   struct ext3_new_group_data *input)
 168 {
 169         struct ext3_sb_info *sbi = EXT3_SB(sb);
 170         unsigned long start = input->group * sbi->s_blocks_per_group +
 171                 le32_to_cpu(sbi->s_es->s_first_data_block);
 172         int reserved_gdb = ext3_bg_has_super(sb, input->group) ?
 173                 le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks) : 0;
 174         unsigned long gdblocks = ext3_bg_num_gdb(sb, input->group);
 175         struct buffer_head *bh;
 176         handle_t *handle;
 177         unsigned long block;
 178         ext3_grpblk_t bit;
 179         int i;
 180         int err = 0, err2;
 181
 182         handle = ext3_journal_start_sb(sb, reserved_gdb + gdblocks +
 183                                        2 + sbi->s_itb_per_group);
 184         if (IS_ERR(handle))
 185                 return PTR_ERR(handle);
 186
 187         lock_super(sb);
 188         if (input->group != sbi->s_groups_count) {
 189                 err = -EBUSY;
 190                 goto exit_journal;
 191         }
 192
 193         if (IS_ERR(bh = bclean(handle, sb, input->block_bitmap))) {
 194                 err = PTR_ERR(bh);
 195                 goto exit_journal;
 196         }
 197
 198         if (ext3_bg_has_super(sb, input->group)) {
 199                 ext3_debug("mark backup superblock %#04lx (+0)\n", start);
 200                 ext3_set_bit(0, bh->b_data);
 201         }
 202
 203         /* Copy all of the GDT blocks into the backup in this group */
 204         for (i = 0, bit = 1, block = start + 1;
 205              i < gdblocks; i++, block++, bit++) {
 206                 struct buffer_head *gdb;
 207
 208                 ext3_debug("update backup group %#04lx (+%d)\n", block, bit);
 209
 210                 gdb = sb_getblk(sb, block);
 211                 if (!gdb) {
 212                         err = -EIO;
 213                         goto exit_bh;
 214                 }
 215                 if ((err = ext3_journal_get_write_access(handle, gdb))) {
 216                         brelse(gdb);
 217                         goto exit_bh;
 218                 }
 219                 lock_buffer(bh);
 220                 memcpy(gdb->b_data, sbi->s_group_desc[i]->b_data, bh->b_size);
 221                 set_buffer_uptodate(gdb);
 222                 unlock_buffer(bh);
 223                 ext3_journal_dirty_metadata(handle, gdb);
 224                 ext3_set_bit(bit, bh->b_data);
 225                 brelse(gdb);
 226         }
 227
 228         /* Zero out all of the reserved backup group descriptor table blocks */
 229         for (i = 0, bit = gdblocks + 1, block = start + bit;
 230              i < reserved_gdb; i++, block++, bit++) {
 231                 struct buffer_head *gdb;
 232
 233                 ext3_debug("clear reserved block %#04lx (+%d)\n", block, bit);
 234
 235                 if (IS_ERR(gdb = bclean(handle, sb, block))) {
 236                         err = PTR_ERR(bh);
 237                         goto exit_bh;
 238                 }
 239                 ext3_journal_dirty_metadata(handle, gdb);
 240                 ext3_set_bit(bit, bh->b_data);
 241                 brelse(gdb);
 242         }
 243         ext3_debug("mark block bitmap %#04x (+%ld)\n", input->block_bitmap,
 244                    input->block_bitmap - start);
 245         ext3_set_bit(input->block_bitmap - start, bh->b_data);
 246         ext3_debug("mark inode bitmap %#04x (+%ld)\n", input->inode_bitmap,
 247                    input->inode_bitmap - start);
 248         ext3_set_bit(input->inode_bitmap - start, bh->b_data);
 249
 250         /* Zero out all of the inode table blocks */
 251         for (i = 0, block = input->inode_table, bit = block - start;
 252              i < sbi->s_itb_per_group; i++, bit++, block++) {
 253                 struct buffer_head *it;
 254
 255                 ext3_debug("clear inode block %#04lx (+%d)\n", block, bit);
 256                 if (IS_ERR(it = bclean(handle, sb, block))) {
 257                         err = PTR_ERR(it);
 258                         goto exit_bh;
 259                 }
 260                 ext3_journal_dirty_metadata(handle, it);
 261                 brelse(it);
 262                 ext3_set_bit(bit, bh->b_data);
 263         }
 264         mark_bitmap_end(input->blocks_count, EXT3_BLOCKS_PER_GROUP(sb),
 265                         bh->b_data);
 266         ext3_journal_dirty_metadata(handle, bh);
 267         brelse(bh);
 268
 269         /* Mark unused entries in inode bitmap used */
 270         ext3_debug("clear inode bitmap %#04x (+%ld)\n",
 271                    input->inode_bitmap, input->inode_bitmap - start);
 272         if (IS_ERR(bh = bclean(handle, sb, input->inode_bitmap))) {
 273                 err = PTR_ERR(bh);
 274                 goto exit_journal;
 275         }
 276
 277         mark_bitmap_end(EXT3_INODES_PER_GROUP(sb), EXT3_BLOCKS_PER_GROUP(sb),
 278                         bh->b_data);
 279         ext3_journal_dirty_metadata(handle, bh);
 280 exit_bh:
 281         brelse(bh);
 282
 283 exit_journal:
 284         unlock_super(sb);
 285         if ((err2 = ext3_journal_stop(handle)) && !err)
 286                 err = err2;
 287
 288         return err;
 289 }
 290
 291 /*
 292  * Iterate through the groups which hold BACKUP superblock/GDT copies in an
 293  * ext3 filesystem.  The counters should be initialized to 1, 5, and 7 before
 294  * calling this for the first time.  In a sparse filesystem it will be the
 295  * sequence of powers of 3, 5, and 7: 1, 3, 5, 7, 9, 25, 27, 49, 81, ...
 296  * For a non-sparse filesystem it will be every group: 1, 2, 3, 4, ...
 297  */
 298 static unsigned ext3_list_backups(struct super_block *sb, unsigned *three,
 299                                   unsigned *five, unsigned *seven)
 300 {
 301         unsigned *min = three;
 302         int mult = 3;
 303         unsigned ret;
 304
 305         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
 306                                         EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
 307                 ret = *min;
 308                 *min += 1;
 309                 return ret;
 310         }
 311
 312         if (*five < *min) {
 313                 min = five;
 314                 mult = 5;
 315         }
 316         if (*seven < *min) {
 317                 min = seven;
 318                 mult = 7;
 319         }
 320
 321         ret = *min;
 322         *min *= mult;
 323
 324         return ret;
 325 }
 326
 327 /*
 328  * Check that all of the backup GDT blocks are held in the primary GDT block.
 329  * It is assumed that they are stored in group order.  Returns the number of
 330  * groups in current filesystem that have BACKUPS, or -ve error code.
 331  */
 332 static int verify_reserved_gdb(struct super_block *sb,
 333                                struct buffer_head *primary)
 334 {
 335         const unsigned long blk = primary->b_blocknr;
 336         const unsigned long end = EXT3_SB(sb)->s_groups_count;
 337         unsigned three = 1;
 338         unsigned five = 5;
 339         unsigned seven = 7;
 340         unsigned grp;
 341         __u32 *p = (__u32 *)primary->b_data;
 342         int gdbackups = 0;
 343
 344         while ((grp = ext3_list_backups(sb, &three, &five, &seven)) < end) {
 345                 if (le32_to_cpu(*p++) != grp * EXT3_BLOCKS_PER_GROUP(sb) + blk){
 346                         ext3_warning(sb, __FUNCTION__,
 347                                      "reserved GDT %lu missing grp %d (%lu)",
 348                                      blk, grp,
 349                                      grp * EXT3_BLOCKS_PER_GROUP(sb) + blk);
 350                         return -EINVAL;
 351                 }
 352                 if (++gdbackups > EXT3_ADDR_PER_BLOCK(sb))
 353                         return -EFBIG;
 354         }
 355
 356         return gdbackups;
 357 }
 358
 359 /*
 360  * Called when we need to bring a reserved group descriptor table block into
 361  * use from the resize inode.  The primary copy of the new GDT block currently
 362  * is an indirect block (under the double indirect block in the resize inode).
 363  * The new backup GDT blocks will be stored as leaf blocks in this indirect
 364  * block, in group order.  Even though we know all the block numbers we need,
 365  * we check to ensure that the resize inode has actually reserved these blocks.
 366  *
 367  * Don't need to update the block bitmaps because the blocks are still in use.
 368  *
 369  * We get all of the error cases out of the way, so that we are sure to not
 370  * fail once we start modifying the data on disk, because JBD has no rollback.
 371  */
 372 static int add_new_gdb(handle_t *handle, struct inode *inode,
 373                        struct ext3_new_group_data *input,
 374                        struct buffer_head **primary)
 375 {
 376         struct super_block *sb = inode->i_sb;
 377         struct ext3_super_block *es = EXT3_SB(sb)->s_es;
 378         unsigned long gdb_num = input->group / EXT3_DESC_PER_BLOCK(sb);
 379         unsigned long gdblock = EXT3_SB(sb)->s_sbh->b_blocknr + 1 + gdb_num;
 380         struct buffer_head **o_group_desc, **n_group_desc;
 381         struct buffer_head *dind;
 382         int gdbackups;
 383         struct ext3_iloc iloc;
 384         __u32 *data;
 385         int err;
 386
 387         if (test_opt(sb, DEBUG))
 388                 printk(KERN_DEBUG
 389                        "EXT3-fs: ext3_add_new_gdb: adding group block %lu\n",
 390                        gdb_num);
 391
 392         /*
 393          * If we are not using the primary superblock/GDT copy don't resize,
 394          * because the user tools have no way of handling this.  Probably a
 395          * bad time to do it anyways.
 396          */
 397         if (EXT3_SB(sb)->s_sbh->b_blocknr !=
 398             le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) {
 399                 ext3_warning(sb, __FUNCTION__,
 400                         "won't resize using backup superblock at %llu",
 401                         (unsigned long long)EXT3_SB(sb)->s_sbh->b_blocknr);
 402                 return -EPERM;
 403         }
 404
 405         *primary = sb_bread(sb, gdblock);
 406         if (!*primary)
 407                 return -EIO;
 408
 409         if ((gdbackups = verify_reserved_gdb(sb, *primary)) < 0) {
 410                 err = gdbackups;
 411                 goto exit_bh;
 412         }
 413
 414         data = EXT3_I(inode)->i_data + EXT3_DIND_BLOCK;
 415         dind = sb_bread(sb, le32_to_cpu(*data));
 416         if (!dind) {
 417                 err = -EIO;
 418                 goto exit_bh;
 419         }
 420
 421         data = (__u32 *)dind->b_data;
 422         if (le32_to_cpu(data[gdb_num % EXT3_ADDR_PER_BLOCK(sb)]) != gdblock) {
 423                 ext3_warning(sb, __FUNCTION__,
 424                              "new group %u GDT block %lu not reserved",
 425                              input->group, gdblock);
 426                 err = -EINVAL;
 427                 goto exit_dind;
 428         }
 429
 430         if ((err = ext3_journal_get_write_access(handle, EXT3_SB(sb)->s_sbh)))
 431                 goto exit_dind;
 432
 433         if ((err = ext3_journal_get_write_access(handle, *primary)))
 434                 goto exit_sbh;
 435
 436         if ((err = ext3_journal_get_write_access(handle, dind)))
 437                 goto exit_primary;
 438
 439         /* ext3_reserve_inode_write() gets a reference on the iloc */
 440         if ((err = ext3_reserve_inode_write(handle, inode, &iloc)))
 441                 goto exit_dindj;
 442
 443         n_group_desc = (struct buffer_head **)kmalloc((gdb_num + 1) *
 444                                 sizeof(struct buffer_head *), GFP_KERNEL);
 445         if (!n_group_desc) {
 446                 err = -ENOMEM;
 447                 ext3_warning (sb, __FUNCTION__,
 448                               "not enough memory for %lu groups", gdb_num + 1);
 449                 goto exit_inode;
 450         }
 451
 452         /*
 453          * Finally, we have all of the possible failures behind us...
 454          *
 455          * Remove new GDT block from inode double-indirect block and clear out
 456          * the new GDT block for use (which also "frees" the backup GDT blocks
 457          * from the reserved inode).  We don't need to change the bitmaps for
 458          * these blocks, because they are marked as in-use from being in the
 459          * reserved inode, and will become GDT blocks (primary and backup).
 460          */
 461         data[gdb_num % EXT3_ADDR_PER_BLOCK(sb)] = 0;
 462         ext3_journal_dirty_metadata(handle, dind);
 463         brelse(dind);
 464         inode->i_blocks -= (gdbackups + 1) * sb->s_blocksize >> 9;
 465         ext3_mark_iloc_dirty(handle, inode, &iloc);
 466         memset((*primary)->b_data, 0, sb->s_blocksize);
 467         ext3_journal_dirty_metadata(handle, *primary);
 468
 469         o_group_desc = EXT3_SB(sb)->s_group_desc;
 470         memcpy(n_group_desc, o_group_desc,
 471                EXT3_SB(sb)->s_gdb_count * sizeof(struct buffer_head *));
 472         n_group_desc[gdb_num] = *primary;
 473         EXT3_SB(sb)->s_group_desc = n_group_desc;
 474         EXT3_SB(sb)->s_gdb_count++;
 475         kfree(o_group_desc);
 476
 477         es->s_reserved_gdt_blocks =
 478                 cpu_to_le16(le16_to_cpu(es->s_reserved_gdt_blocks) - 1);
 479         ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
 480
 481         return 0;
 482
 483 exit_inode:
 484         //ext3_journal_release_buffer(handle, iloc.bh);
 485         brelse(iloc.bh);
 486 exit_dindj:
 487         //ext3_journal_release_buffer(handle, dind);
 488 exit_primary:
 489         //ext3_journal_release_buffer(handle, *primary);
 490 exit_sbh:
 491         //ext3_journal_release_buffer(handle, *primary);
 492 exit_dind:
 493         brelse(dind);
 494 exit_bh:
 495         brelse(*primary);
 496
 497         ext3_debug("leaving with error %d\n", err);
 498         return err;
 499 }
 500
 501 /*
 502  * Called when we are adding a new group which has a backup copy of each of
 503  * the GDT blocks (i.e. sparse group) and there are reserved GDT blocks.
 504  * We need to add these reserved backup GDT blocks to the resize inode, so
 505  * that they are kept for future resizing and not allocated to files.
 506  *
 507  * Each reserved backup GDT block will go into a different indirect block.
 508  * The indirect blocks are actually the primary reserved GDT blocks,
 509  * so we know in advance what their block numbers are.  We only get the
 510  * double-indirect block to verify it is pointing to the primary reserved
 511  * GDT blocks so we don't overwrite a data block by accident.  The reserved
 512  * backup GDT blocks are stored in their reserved primary GDT block.
 513  */
 514 static int reserve_backup_gdb(handle_t *handle, struct inode *inode,
 515                               struct ext3_new_group_data *input)
 516 {
 517         struct super_block *sb = inode->i_sb;
 518         int reserved_gdb =le16_to_cpu(EXT3_SB(sb)->s_es->s_reserved_gdt_blocks);
 519         struct buffer_head **primary;
 520         struct buffer_head *dind;
 521         struct ext3_iloc iloc;
 522         unsigned long blk;
 523         __u32 *data, *end;
 524         int gdbackups = 0;
 525         int res, i;
 526         int err;
 527
 528         primary = kmalloc(reserved_gdb * sizeof(*primary), GFP_KERNEL);
 529         if (!primary)
 530                 return -ENOMEM;
 531
 532         data = EXT3_I(inode)->i_data + EXT3_DIND_BLOCK;
 533         dind = sb_bread(sb, le32_to_cpu(*data));
 534         if (!dind) {
 535                 err = -EIO;
 536                 goto exit_free;
 537         }
 538
 539         blk = EXT3_SB(sb)->s_sbh->b_blocknr + 1 + EXT3_SB(sb)->s_gdb_count;
 540         data = (__u32 *)dind->b_data + EXT3_SB(sb)->s_gdb_count;
 541         end = (__u32 *)dind->b_data + EXT3_ADDR_PER_BLOCK(sb);
 542
 543         /* Get each reserved primary GDT block and verify it holds backups */
 544         for (res = 0; res < reserved_gdb; res++, blk++) {
 545                 if (le32_to_cpu(*data) != blk) {
 546                         ext3_warning(sb, __FUNCTION__,
 547                                      "reserved block %lu not at offset %ld",
 548                                      blk, (long)(data - (__u32 *)dind->b_data));
 549                         err = -EINVAL;
 550                         goto exit_bh;
 551                 }
 552                 primary[res] = sb_bread(sb, blk);
 553                 if (!primary[res]) {
 554                         err = -EIO;
 555                         goto exit_bh;
 556                 }
 557                 if ((gdbackups = verify_reserved_gdb(sb, primary[res])) < 0) {
 558                         brelse(primary[res]);
 559                         err = gdbackups;
 560                         goto exit_bh;
 561                 }
 562                 if (++data >= end)
 563                         data = (__u32 *)dind->b_data;
 564         }
 565
 566         for (i = 0; i < reserved_gdb; i++) {
 567                 if ((err = ext3_journal_get_write_access(handle, primary[i]))) {
 568                         /*
 569                         int j;
 570                         for (j = 0; j < i; j++)
 571                                 ext3_journal_release_buffer(handle, primary[j]);
 572                          */
 573                         goto exit_bh;
 574                 }
 575         }
 576
 577         if ((err = ext3_reserve_inode_write(handle, inode, &iloc)))
 578                 goto exit_bh;
 579
 580         /*
 581          * Finally we can add each of the reserved backup GDT blocks from
 582          * the new group to its reserved primary GDT block.
 583          */
 584         blk = input->group * EXT3_BLOCKS_PER_GROUP(sb);
 585         for (i = 0; i < reserved_gdb; i++) {
 586                 int err2;
 587                 data = (__u32 *)primary[i]->b_data;
 588                 /* printk("reserving backup %lu[%u] = %lu\n",
 589                        primary[i]->b_blocknr, gdbackups,
 590                        blk + primary[i]->b_blocknr); */
 591                 data[gdbackups] = cpu_to_le32(blk + primary[i]->b_blocknr);
 592                 err2 = ext3_journal_dirty_metadata(handle, primary[i]);
 593                 if (!err)
 594                         err = err2;
 595         }
 596         inode->i_blocks += reserved_gdb * sb->s_blocksize >> 9;
 597         ext3_mark_iloc_dirty(handle, inode, &iloc);
 598
 599 exit_bh:
 600         while (--res >= 0)
 601                 brelse(primary[res]);
 602         brelse(dind);
 603
 604 exit_free:
 605         kfree(primary);
 606
 607         return err;
 608 }
 609
 610 /*
 611  * Update the backup copies of the ext3 metadata.  These don't need to be part
 612  * of the main resize transaction, because e2fsck will re-write them if there
 613  * is a problem (basically only OOM will cause a problem).  However, we
 614  * _should_ update the backups if possible, in case the primary gets trashed
 615  * for some reason and we need to run e2fsck from a backup superblock.  The
 616  * important part is that the new block and inode counts are in the backup
 617  * superblocks, and the location of the new group metadata in the GDT backups.
 618  *
 619  * We do not need lock_super() for this, because these blocks are not
 620  * otherwise touched by the filesystem code when it is mounted.  We don't
 621  * need to worry about last changing from sbi->s_groups_count, because the
 622  * worst that can happen is that we do not copy the full number of backups
 623  * at this time.  The resize which changed s_groups_count will backup again.
 624  */
 625 static void update_backups(struct super_block *sb,
 626                            int blk_off, char *data, int size)
 627 {
 628         struct ext3_sb_info *sbi = EXT3_SB(sb);
 629         const unsigned long last = sbi->s_groups_count;
 630         const int bpg = EXT3_BLOCKS_PER_GROUP(sb);
 631         unsigned three = 1;
 632         unsigned five = 5;
 633         unsigned seven = 7;
 634         unsigned group;
 635         int rest = sb->s_blocksize - size;
 636         handle_t *handle;
 637         int err = 0, err2;
 638
 639         handle = ext3_journal_start_sb(sb, EXT3_MAX_TRANS_DATA);
 640         if (IS_ERR(handle)) {
 641                 group = 1;
 642                 err = PTR_ERR(handle);
 643                 goto exit_err;
 644         }
 645
 646         while ((group = ext3_list_backups(sb, &three, &five, &seven)) < last) {
 647                 struct buffer_head *bh;
 648
 649                 /* Out of journal space, and can't get more - abort - so sad */
 650                 if (handle->h_buffer_credits == 0 &&
 651                     ext3_journal_extend(handle, EXT3_MAX_TRANS_DATA) &&
 652                     (err = ext3_journal_restart(handle, EXT3_MAX_TRANS_DATA)))
 653                         break;
 654
 655                 bh = sb_getblk(sb, group * bpg + blk_off);
 656                 if (!bh) {
 657                         err = -EIO;
 658                         break;
 659                 }
 660                 ext3_debug("update metadata backup %#04lx\n",
 661                           (unsigned long)bh->b_blocknr);
 662                 if ((err = ext3_journal_get_write_access(handle, bh)))
 663                         break;
 664                 lock_buffer(bh);
 665                 memcpy(bh->b_data, data, size);
 666                 if (rest)
 667                         memset(bh->b_data + size, 0, rest);
 668                 set_buffer_uptodate(bh);
 669                 unlock_buffer(bh);
 670                 ext3_journal_dirty_metadata(handle, bh);
 671                 brelse(bh);
 672         }
 673         if ((err2 = ext3_journal_stop(handle)) && !err)
 674                 err = err2;
 675
 676         /*
 677          * Ugh! Need to have e2fsck write the backup copies.  It is too
 678          * late to revert the resize, we shouldn't fail just because of
 679          * the backup copies (they are only needed in case of corruption).
 680          *
 681          * However, if we got here we have a journal problem too, so we
 682          * can't really start a transaction to mark the superblock.
 683          * Chicken out and just set the flag on the hope it will be written
 684          * to disk, and if not - we will simply wait until next fsck.
 685          */
 686 exit_err:
 687         if (err) {
 688                 ext3_warning(sb, __FUNCTION__,
 689                              "can't update backup for group %d (err %d), "
 690                              "forcing fsck on next reboot", group, err);
 691                 sbi->s_mount_state &= ~EXT3_VALID_FS;
 692                 sbi->s_es->s_state &= ~cpu_to_le16(EXT3_VALID_FS);
 693                 mark_buffer_dirty(sbi->s_sbh);
 694         }
 695 }
 696
 697 /* Add group descriptor data to an existing or new group descriptor block.
 698  * Ensure we handle all possible error conditions _before_ we start modifying
 699  * the filesystem, because we cannot abort the transaction and not have it
 700  * write the data to disk.
 701  *
 702  * If we are on a GDT block boundary, we need to get the reserved GDT block.
 703  * Otherwise, we may need to add backup GDT blocks for a sparse group.
 704  *
 705  * We only need to hold the superblock lock while we are actually adding
 706  * in the new group's counts to the superblock.  Prior to that we have
 707  * not really "added" the group at all.  We re-check that we are still
 708  * adding in the last group in case things have changed since verifying.
 709  */
 710 int ext3_group_add(struct super_block *sb, struct ext3_new_group_data *input)
 711 {
 712         struct ext3_sb_info *sbi = EXT3_SB(sb);
 713         struct ext3_super_block *es = sbi->s_es;
 714         int reserved_gdb = ext3_bg_has_super(sb, input->group) ?
 715                 le16_to_cpu(es->s_reserved_gdt_blocks) : 0;
 716         struct buffer_head *primary = NULL;
 717         struct ext3_group_desc *gdp;
 718         struct inode *inode = NULL;
 719         handle_t *handle;
 720         int gdb_off, gdb_num;
 721         int err, err2;
 722
 723         gdb_num = input->group / EXT3_DESC_PER_BLOCK(sb);
 724         gdb_off = input->group % EXT3_DESC_PER_BLOCK(sb);
 725
 726         if (gdb_off == 0 && !EXT3_HAS_RO_COMPAT_FEATURE(sb,
 727                                         EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
 728                 ext3_warning(sb, __FUNCTION__,
 729                              "Can't resize non-sparse filesystem further");
 730                 return -EPERM;
 731         }
 732
 733         if (reserved_gdb || gdb_off == 0) {
 734                 if (!EXT3_HAS_COMPAT_FEATURE(sb,
 735                                              EXT3_FEATURE_COMPAT_RESIZE_INODE)){
 736                         ext3_warning(sb, __FUNCTION__,
 737                                      "No reserved GDT blocks, can't resize");
 738                         return -EPERM;
 739                 }
 740                 inode = iget(sb, EXT3_RESIZE_INO);
 741                 if (!inode || is_bad_inode(inode)) {
 742                         ext3_warning(sb, __FUNCTION__,
 743                                      "Error opening resize inode");
 744                         iput(inode);
 745                         return -ENOENT;
 746                 }
 747         }
 748
 749         if ((err = verify_group_input(sb, input)))
 750                 goto exit_put;
 751
 752         if ((err = setup_new_group_blocks(sb, input)))
 753                 goto exit_put;
 754
 755         /*
 756          * We will always be modifying at least the superblock and a GDT
 757          * block.  If we are adding a group past the last current GDT block,
 758          * we will also modify the inode and the dindirect block.  If we
 759          * are adding a group with superblock/GDT backups  we will also
 760          * modify each of the reserved GDT dindirect blocks.
 761          */
 762         handle = ext3_journal_start_sb(sb,
 763                                        ext3_bg_has_super(sb, input->group) ?
 764                                        3 + reserved_gdb : 4);
 765         if (IS_ERR(handle)) {
 766                 err = PTR_ERR(handle);
 767                 goto exit_put;
 768         }
 769
 770         lock_super(sb);
 771         if (input->group != sbi->s_groups_count) {
 772                 ext3_warning(sb, __FUNCTION__,
 773                              "multiple resizers run on filesystem!");
 774                 err = -EBUSY;
 775                 goto exit_journal;
 776         }
 777
 778         if ((err = ext3_journal_get_write_access(handle, sbi->s_sbh)))
 779                 goto exit_journal;
 780
 781         /*
 782          * We will only either add reserved group blocks to a backup group
 783          * or remove reserved blocks for the first group in a new group block.
 784          * Doing both would be mean more complex code, and sane people don't
 785          * use non-sparse filesystems anymore.  This is already checked above.
 786          */
 787         if (gdb_off) {
 788                 primary = sbi->s_group_desc[gdb_num];
 789                 if ((err = ext3_journal_get_write_access(handle, primary)))
 790                         goto exit_journal;
 791
 792                 if (reserved_gdb && ext3_bg_num_gdb(sb, input->group) &&
 793                     (err = reserve_backup_gdb(handle, inode, input)))
 794                         goto exit_journal;
 795         } else if ((err = add_new_gdb(handle, inode, input, &primary)))
 796                 goto exit_journal;
 797
 798         /*
 799          * OK, now we've set up the new group.  Time to make it active.
 800          *
 801          * Current kernels don't lock all allocations via lock_super(),
 802          * so we have to be safe wrt. concurrent accesses the group
 803          * data.  So we need to be careful to set all of the relevant
 804          * group descriptor data etc. *before* we enable the group.
 805          *
 806          * The key field here is sbi->s_groups_count: as long as
 807          * that retains its old value, nobody is going to access the new
 808          * group.
 809          *
 810          * So first we update all the descriptor metadata for the new
 811          * group; then we update the total disk blocks count; then we
 812          * update the groups count to enable the group; then finally we
 813          * update the free space counts so that the system can start
 814          * using the new disk blocks.
 815          */
 816
 817         /* Update group descriptor block for new group */
 818         gdp = (struct ext3_group_desc *)primary->b_data + gdb_off;
 819
 820         gdp->bg_block_bitmap = cpu_to_le32(input->block_bitmap);
 821         gdp->bg_inode_bitmap = cpu_to_le32(input->inode_bitmap);
 822         gdp->bg_inode_table = cpu_to_le32(input->inode_table);
 823         gdp->bg_free_blocks_count = cpu_to_le16(input->free_blocks_count);
 824         gdp->bg_free_inodes_count = cpu_to_le16(EXT3_INODES_PER_GROUP(sb));
 825
 826         /*
 827          * Make the new blocks and inodes valid next.  We do this before
 828          * increasing the group count so that once the group is enabled,
 829          * all of its blocks and inodes are already valid.
 830          *
 831          * We always allocate group-by-group, then block-by-block or
 832          * inode-by-inode within a group, so enabling these
 833          * blocks/inodes before the group is live won't actually let us
 834          * allocate the new space yet.
 835          */
 836         es->s_blocks_count = cpu_to_le32(le32_to_cpu(es->s_blocks_count) +
 837                 input->blocks_count);
 838         es->s_inodes_count = cpu_to_le32(le32_to_cpu(es->s_inodes_count) +
 839                 EXT3_INODES_PER_GROUP(sb));
 840
 841         /*
 842          * We need to protect s_groups_count against other CPUs seeing
 843          * inconsistent state in the superblock.
 844          *
 845          * The precise rules we use are:
 846          *
 847          * * Writers of s_groups_count *must* hold lock_super
 848          * AND
 849          * * Writers must perform a smp_wmb() after updating all dependent
 850          *   data and before modifying the groups count
 851          *
 852          * * Readers must hold lock_super() over the access
 853          * OR
 854          * * Readers must perform an smp_rmb() after reading the groups count
 855          *   and before reading any dependent data.
 856          *
 857          * NB. These rules can be relaxed when checking the group count
 858          * while freeing data, as we can only allocate from a block
 859          * group after serialising against the group count, and we can
 860          * only then free after serialising in turn against that
 861          * allocation.
 862          */
 863         smp_wmb();
 864
 865         /* Update the global fs size fields */
 866         sbi->s_groups_count++;
 867
 868         ext3_journal_dirty_metadata(handle, primary);
 869
 870         /* Update the reserved block counts only once the new group is
 871          * active. */
 872         es->s_r_blocks_count = cpu_to_le32(le32_to_cpu(es->s_r_blocks_count) +
 873                 input->reserved_blocks);
 874
 875         /* Update the free space counts */
 876         percpu_counter_mod(&sbi->s_freeblocks_counter,
 877                            input->free_blocks_count);
 878         percpu_counter_mod(&sbi->s_freeinodes_counter,
 879                            EXT3_INODES_PER_GROUP(sb));
 880
 881         ext3_journal_dirty_metadata(handle, sbi->s_sbh);
 882         sb->s_dirt = 1;
 883
 884 exit_journal:
 885         unlock_super(sb);
 886         if ((err2 = ext3_journal_stop(handle)) && !err)
 887                 err = err2;
 888         if (!err) {
 889                 update_backups(sb, sbi->s_sbh->b_blocknr, (char *)es,
 890                                sizeof(struct ext3_super_block));
 891                 update_backups(sb, primary->b_blocknr, primary->b_data,
 892                                primary->b_size);
 893         }
 894 exit_put:
 895         iput(inode);
 896         return err;
 897 } /* ext3_group_add */
 898
 899 /* Extend the filesystem to the new number of blocks specified.  This entry
 900  * point is only used to extend the current filesystem to the end of the last
 901  * existing group.  It can be accessed via ioctl, or by "remount,resize=<size>"
 902  * for emergencies (because it has no dependencies on reserved blocks).
 903  *
 904  * If we _really_ wanted, we could use default values to call ext3_group_add()
 905  * allow the "remount" trick to work for arbitrary resizing, assuming enough
 906  * GDT blocks are reserved to grow to the desired size.
 907  */
 908 int ext3_group_extend(struct super_block *sb, struct ext3_super_block *es,
 909                       unsigned long n_blocks_count)
 910 {
 911         unsigned long o_blocks_count;
 912         unsigned long o_groups_count;
 913         ext3_grpblk_t last;
 914         ext3_grpblk_t add;
 915         struct buffer_head * bh;
 916         handle_t *handle;
 917         int err;
 918         unsigned long freed_blocks;
 919
 920         /* We don't need to worry about locking wrt other resizers just
 921          * yet: we're going to revalidate es->s_blocks_count after
 922          * taking lock_super() below. */
 923         o_blocks_count = le32_to_cpu(es->s_blocks_count);
 924         o_groups_count = EXT3_SB(sb)->s_groups_count;
 925
 926         if (test_opt(sb, DEBUG))
 927                 printk(KERN_DEBUG "EXT3-fs: extending last group from %lu to %lu blocks\n",
 928                        o_blocks_count, n_blocks_count);
 929
 930         if (n_blocks_count == 0 || n_blocks_count == o_blocks_count)
 931                 return 0;
 932
 933         if (n_blocks_count > (sector_t)(~0ULL) >> (sb->s_blocksize_bits - 9)) {
 934                 printk(KERN_ERR "EXT3-fs: filesystem on %s:"
 935                         " too large to resize to %lu blocks safely\n",
 936                         sb->s_id, n_blocks_count);
 937                 if (sizeof(sector_t) < 8)
 938                         ext3_warning(sb, __FUNCTION__,
 939                         "CONFIG_LBD not enabled\n");
 940                 return -EINVAL;
 941         }
 942
 943         if (n_blocks_count < o_blocks_count) {
 944                 ext3_warning(sb, __FUNCTION__,
 945                              "can't shrink FS - resize aborted");
 946                 return -EBUSY;
 947         }
 948
 949         /* Handle the remaining blocks in the last group only. */
 950         last = (o_blocks_count - le32_to_cpu(es->s_first_data_block)) %
 951                 EXT3_BLOCKS_PER_GROUP(sb);
 952
 953         if (last == 0) {
 954                 ext3_warning(sb, __FUNCTION__,
 955                              "need to use ext2online to resize further");
 956                 return -EPERM;
 957         }
 958
 959         add = EXT3_BLOCKS_PER_GROUP(sb) - last;
 960
 961         if (o_blocks_count + add > n_blocks_count)
 962                 add = n_blocks_count - o_blocks_count;
 963
 964         if (o_blocks_count + add < n_blocks_count)
 965                 ext3_warning(sb, __FUNCTION__,
 966                              "will only finish group (%lu blocks, %u new)",
 967                              o_blocks_count + add, add);
 968
 969         /* See if the device is actually as big as what was requested */
 970         bh = sb_bread(sb, o_blocks_count + add -1);
 971         if (!bh) {
 972                 ext3_warning(sb, __FUNCTION__,
 973                              "can't read last block, resize aborted");
 974                 return -ENOSPC;
 975         }
 976         brelse(bh);
 977
 978         /* We will update the superblock, one block bitmap, and
 979          * one group descriptor via ext3_free_blocks().
 980          */
 981         handle = ext3_journal_start_sb(sb, 3);
 982         if (IS_ERR(handle)) {
 983                 err = PTR_ERR(handle);
 984                 ext3_warning(sb, __FUNCTION__, "error %d on journal start",err);
 985                 goto exit_put;
 986         }
 987
 988         lock_super(sb);
 989         if (o_blocks_count != le32_to_cpu(es->s_blocks_count)) {
 990                 ext3_warning(sb, __FUNCTION__,
 991                              "multiple resizers run on filesystem!");
 992                 unlock_super(sb);
 993                 err = -EBUSY;
 994                 goto exit_put;
 995         }
 996
 997         if ((err = ext3_journal_get_write_access(handle,
 998                                                  EXT3_SB(sb)->s_sbh))) {
 999                 ext3_warning(sb, __FUNCTION__,
1000                              "error %d on journal write access", err);
1001                 unlock_super(sb);
1002                 ext3_journal_stop(handle);
1003                 goto exit_put;
1004         }
1005         es->s_blocks_count = cpu_to_le32(o_blocks_count + add);
1006         ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
1007         sb->s_dirt = 1;
1008         unlock_super(sb);
1009         ext3_debug("freeing blocks %lu through %lu\n", o_blocks_count,
1010                    o_blocks_count + add);
1011         ext3_free_blocks_sb(handle, sb, o_blocks_count, add, &freed_blocks);
1012         ext3_debug("freed blocks %lu through %lu\n", o_blocks_count,
1013                    o_blocks_count + add);
1014         if ((err = ext3_journal_stop(handle)))
1015                 goto exit_put;
1016         if (test_opt(sb, DEBUG))
1017                 printk(KERN_DEBUG "EXT3-fs: extended group to %u blocks\n",
1018                        le32_to_cpu(es->s_blocks_count));
1019         update_backups(sb, EXT3_SB(sb)->s_sbh->b_blocknr, (char *)es,
1020                        sizeof(struct ext3_super_block));
1021 exit_put:
1022         return err;
1023 } /* ext3_group_extend */